diff --git a/Dockerfile.voice b/Dockerfile.voice
new file mode 100644
index 0000000..b8af0eb
--- /dev/null
+++ b/Dockerfile.voice
@@ -0,0 +1,13 @@
+FROM python:3.12-slim
+
+WORKDIR /app
+COPY pyproject.toml README.md* ./
+COPY gogo ./gogo
+COPY alembic ./alembic
+COPY alembic.ini ./
+
+# voice extra pulls faster-whisper (CPU by default; CUDA wheels on the GPU node)
+RUN pip install --no-cache-dir -e '.[voice]'
+
+EXPOSE 9092
+CMD ["python", "-m", "gogo.voice.server"]
diff --git a/alembic/versions/9dce40d5c9ce_call_registrations.py b/alembic/versions/9dce40d5c9ce_call_registrations.py
new file mode 100644
index 0000000..a0c12ac
--- /dev/null
+++ b/alembic/versions/9dce40d5c9ce_call_registrations.py
@@ -0,0 +1,41 @@
+"""call registrations
+
+Revision ID: 9dce40d5c9ce
+Revises: 62fcb10345fa
+Create Date: 2026-07-11 10:58:21.464071
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+
+revision: str = '9dce40d5c9ce'
+down_revision: Union[str, None] = '62fcb10345fa'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table('call_registrations',
+    sa.Column('id', sa.Uuid(), nullable=False),
+    sa.Column('tenant_id', sa.Uuid(), nullable=False),
+    sa.Column('caller_msisdn', sa.String(length=30), nullable=False),
+    sa.Column('created_at', sa.DateTime(timezone=True), nullable=False),
+    sa.Column('answered_by_human', sa.Boolean(), nullable=False),
+    sa.Column('handled_by_agent', sa.Boolean(), nullable=False),
+    sa.Column('finalized', sa.Boolean(), nullable=False),
+    sa.ForeignKeyConstraint(['tenant_id'], ['tenants.id'], ),
+    sa.PrimaryKeyConstraint('id')
+    )
+    op.create_index(op.f('ix_call_registrations_tenant_id'), 'call_registrations', ['tenant_id'], unique=False)
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_index(op.f('ix_call_registrations_tenant_id'), table_name='call_registrations')
+    op.drop_table('call_registrations')
+    # ### end Alembic commands ###
diff --git a/docker-compose.yml b/docker-compose.yml
index 02b5b55..c8ee6f4 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -33,6 +33,37 @@ services:
       sh -c "alembic upgrade head &&
              uvicorn gogo.main:app --host 0.0.0.0 --port 8000"
 
+  # Voice pipeline (M4): AudioSocket server with STT/TTS/LLM.
+  # Heavy (faster-whisper); start with: docker compose --profile voice up
+  voice:
+    build:
+      context: .
+      dockerfile: Dockerfile.voice
+    profiles: ["voice"]
+    depends_on:
+      db:
+        condition: service_healthy
+    environment:
+      GOGO_DATABASE_URL: postgresql+psycopg://gogo:gogo@db:5432/gogo
+      GOGO_RECORDINGS_DIR: /data/recordings
+    env_file:
+      - path: .env
+        required: false
+    ports:
+      - "127.0.0.1:9092:9092"
+    volumes:
+      - recordings:/data/recordings
+
+  # Asterisk (M4): softphones register here; agent calls bridge to `voice`.
+  # Config is generated: python -m gogo.telephony.asterisk --out ./asterisk/conf
+  asterisk:
+    image: andrius/asterisk:alpine-20.5.2
+    profiles: ["voice"]
+    network_mode: host  # SIP/RTP is painful behind NAT; host networking for MVP
+    volumes:
+      - ./asterisk/conf/pjsip.conf:/etc/asterisk/pjsip.conf:ro
+      - ./asterisk/conf/extensions.conf:/etc/asterisk/extensions.conf:ro
+
 volumes:
   pgdata:
   recordings:
diff --git a/docs/VOICE_TESTING.md b/docs/VOICE_TESTING.md
new file mode 100644
index 0000000..b4ee908
--- /dev/null
+++ b/docs/VOICE_TESTING.md
@@ -0,0 +1,69 @@
+# M4 — Software-only voice testing (no GSM hardware)
+
+Everything here runs on a laptop/DC per §16: desktop/mobile softphones register
+directly to Asterisk over IP; the GSM leg is the only thing not exercised
+(that's M5 + the joint hardware session).
+
+## 1. Prereqs
+
+- Postgres running, schema migrated, demo tenant seeded:
+  ```bash
+  .venv/bin/alembic upgrade head
+  .venv/bin/python -m gogo.seed
+  ```
+- `.env` with at least `GOGO_ANTHROPIC_API_KEY` (agent) and one TTS key
+  (`GOGO_AZURE_SPEECH_KEY` + region, or `GOGO_ELEVENLABS_API_KEY`).
+- Whisper model downloads on first run (`GOGO_WHISPER_MODEL=small` is fine on CPU).
+
+## 2. Start the stack
+
+```bash
+# backend API (internal dialplan endpoints live here)
+.venv/bin/uvicorn gogo.main:app --host 0.0.0.0 --port 8000
+
+# voice pipeline (AudioSocket server on :9092)
+.venv/bin/python -m gogo.voice.server
+
+# generate Asterisk config from the DB, then start Asterisk
+.venv/bin/python -m gogo.telephony.asterisk --out ./asterisk/conf \
+    --backend http://127.0.0.1:8000 --voice 127.0.0.1:9092
+docker compose --profile voice up asterisk
+```
+
+## 3. Softphones
+
+Two SIP accounts exist per tenant after config generation:
+
+| Role | Username | Password | Purpose |
+|---|---|---|---|
+| Worker | from dashboard → Telefonija → QR | generated | rings before the agent |
+| Test caller | `test-caller-salon-merima` | `test-<first 10 chars of widget key>` (in pjsip.conf) | simulates the customer |
+
+Register the worker softphone (e.g. Linphone on a phone, scan the QR) and the
+test-caller softphone (e.g. Zoiper on the laptop) against the machine's IP.
+
+## 4. Test scenarios
+
+1. **Ring-group first (§5.2):** from the test-caller dial any number (e.g. `100`).
+   The worker softphone rings for `ring_timeout_s`; answer → normal human call,
+   dashboard logs outcome `javilo se osoblje`, no agent minutes metered.
+2. **Agent fallback:** dial again, don't answer → after the timeout the agent
+   picks up, greets with the recording disclosure, books a request. Check:
+   dashboard → Zahtjevi (new request), Pozivi (transcript + audio player),
+   Potrošnja (minutes metered), owner email (proposal + .ics).
+3. **Out of hours (A.5):** set working hours to closed (dashboard → Salon),
+   dial → agent answers immediately with the out-of-hours greeting.
+4. **Barge-in:** interrupt the agent mid-sentence — playback must stop and the
+   agent must respond to what you said.
+5. **Abandoned call (§5.5):** dial, hang up while the worker phone still rings →
+   `Pozivi` shows `prekinut` and the missed-call SMS appears in the backend log
+   (console SMS provider).
+6. **Latency (§15):** `calls.trace` (call detail JSON) stores per-turn
+   `stt_ms`/`llm_ms` — target < 1500 ms combined before TTS starts.
+
+## 5. Automated coverage
+
+`tests/test_voice.py` covers the same flows without Asterisk: VAD, the full
+call loop against fakes (booking, unclear speech, hangup, inactivity, barge-in),
+the AudioSocket wire protocol over real TCP, the dialplan internal API
+(ring targets, out-of-hours, abandoned + SMS) and config generation.
diff --git a/gogo/api/internal.py b/gogo/api/internal.py
new file mode 100644
index 0000000..1867de3
--- /dev/null
+++ b/gogo/api/internal.py
@@ -0,0 +1,159 @@
+"""Internal API for the Asterisk dialplan (M4) — plain-text responses because
+the dialplan consumes them via CURL()/CUT().
+
+Auth: shared token (GOGO_INTERNAL_TOKEN) as a query param; this API must only
+be reachable on the internal network regardless.
+"""
+
+from __future__ import annotations
+
+import logging
+import uuid
+
+from fastapi import APIRouter, Depends
+from fastapi.responses import PlainTextResponse
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from gogo.config import get_settings
+from gogo.db import get_session
+from gogo.domain import CallOutcome
+from gogo.hours import DEFAULT_WORKING_HOURS, is_open_at
+from gogo.models import Call, CallRegistration, Tenant, Worker, utcnow
+from gogo.sms import send_sms
+from gogo.sms.templates import render_sms
+
+log = logging.getLogger("gogo.internal")
+
+router = APIRouter()
+
+
+def _authorized(token: str) -> bool:
+    return token == get_settings().internal_token
+
+
+@router.get("/internal/calls/register", response_class=PlainTextResponse)
+async def register_call(
+    tenant: str,
+    caller: str = "",
+    token: str = "",
+    session: AsyncSession = Depends(get_session),
+):
+    """Dialplan entry point. Returns 'uuid,ring_timeout,dial_targets'.
+
+    dial_targets is empty when the agent should answer immediately
+    (out of hours, ring-first disabled, or no active workers §5.2).
+    """
+    if not _authorized(token):
+        return PlainTextResponse("forbidden", status_code=403)
+    t = (
+        await session.execute(select(Tenant).where(Tenant.slug == tenant))
+    ).scalar_one_or_none()
+    if t is None or t.status != "active":
+        return PlainTextResponse("unknown-tenant", status_code=404)
+
+    reg = CallRegistration(tenant_id=t.id, caller_msisdn=caller.strip())
+    session.add(reg)
+    await session.commit()
+
+    targets = ""
+    from zoneinfo import ZoneInfo
+
+    open_now = is_open_at(
+        t.working_hours or DEFAULT_WORKING_HOURS, utcnow(), ZoneInfo(t.timezone)
+    )
+    if t.ring_first_enabled and open_now:
+        workers = (
+            (
+                await session.execute(
+                    select(Worker)
+                    .where(Worker.tenant_id == t.id, Worker.active.is_(True))
+                    .order_by(Worker.sip_username)
+                )
+            )
+            .scalars()
+            .all()
+        )
+        if workers:
+            # '&' rings all at once (default §5.2); sequential strategy is a Phase-2 tuning
+            targets = "&".join(f"PJSIP/{w.sip_username}" for w in workers)
+
+    return f"{reg.id},{t.ring_timeout_s},{targets}"
+
+
+@router.get("/internal/calls/{reg_id}/answered", response_class=PlainTextResponse)
+async def call_answered(
+    reg_id: str,
+    token: str = "",
+    duration: int = 0,
+    session: AsyncSession = Depends(get_session),
+):
+    """Ring group answered by a human — log the call, no agent minutes (§12)."""
+    if not _authorized(token):
+        return PlainTextResponse("forbidden", status_code=403)
+    reg = (
+        await session.execute(
+            select(CallRegistration).where(CallRegistration.id == uuid.UUID(reg_id))
+        )
+    ).scalar_one_or_none()
+    if reg is None:
+        return PlainTextResponse("unknown", status_code=404)
+    reg.answered_by_human = True
+    if not reg.finalized:
+        reg.finalized = True
+        session.add(
+            Call(
+                tenant_id=reg.tenant_id,
+                caller_msisdn=reg.caller_msisdn,
+                outcome=CallOutcome.human_answered.value,
+                duration_s=max(0, duration),
+            )
+        )
+    await session.commit()
+    return "ok"
+
+
+@router.get("/internal/calls/{reg_id}/hangup", response_class=PlainTextResponse)
+async def call_hangup(
+    reg_id: str,
+    status: str = "",
+    token: str = "",
+    session: AsyncSession = Depends(get_session),
+):
+    """Dialplan h-extension. If nobody (human or agent) handled the call, the
+    caller abandoned during ringing → log + missed-call SMS (§5.5)."""
+    if not _authorized(token):
+        return PlainTextResponse("forbidden", status_code=403)
+    reg = (
+        await session.execute(
+            select(CallRegistration).where(CallRegistration.id == uuid.UUID(reg_id))
+        )
+    ).scalar_one_or_none()
+    if reg is None:
+        return PlainTextResponse("unknown", status_code=404)
+    if reg.finalized or reg.answered_by_human or reg.handled_by_agent:
+        return "ok"
+
+    reg.finalized = True
+    tenant = (
+        await session.execute(select(Tenant).where(Tenant.id == reg.tenant_id))
+    ).scalar_one()
+    session.add(
+        Call(
+            tenant_id=reg.tenant_id,
+            caller_msisdn=reg.caller_msisdn,
+            outcome=CallOutcome.abandoned.value,
+        )
+    )
+    if reg.caller_msisdn:
+        link = tenant.website or get_settings().base_url
+        await send_sms(
+            session,
+            tenant.id,
+            reg.caller_msisdn,
+            render_sms(tenant, "missed_call", link=link),
+            kind="missed_call",
+        )
+    await session.commit()
+    log.info("abandoned call for %s from %s (status=%s)", tenant.slug, reg.caller_msisdn, status)
+    return "ok"
diff --git a/gogo/config.py b/gogo/config.py
index 06de9a8..5481a15 100644
--- a/gogo/config.py
+++ b/gogo/config.py
@@ -43,6 +43,12 @@ class Settings(BaseSettings):
     google_client_id: str = ""
     google_client_secret: str = ""
 
+    # Voice pipeline (M4)
+    voice_host: str = "0.0.0.0"
+    voice_port: int = 9092
+    recordings_dir: str = "./recordings"
+    internal_token: str = "dev-internal-token"  # dialplan ↔ backend shared secret
+
     # Proposal engine defaults (overridable per tenant)
     proposal_ttl_hours: int = 24
 
diff --git a/gogo/main.py b/gogo/main.py
index 9623db2..ff83b86 100644
--- a/gogo/main.py
+++ b/gogo/main.py
@@ -30,6 +30,7 @@ async def lifespan(app: FastAPI):
 
 def create_app() -> FastAPI:
     from gogo.admin.router import router as admin_router
+    from gogo.api.internal import router as internal_router
     from gogo.chat.router import router as chat_router
     from gogo.dashboard.router import router as dashboard_router
 
@@ -40,6 +41,7 @@ def create_app() -> FastAPI:
     app.include_router(chat_router)
     app.include_router(dashboard_router)
     app.include_router(admin_router)
+    app.include_router(internal_router)
 
     @app.get("/", include_in_schema=False)
     async def root():
diff --git a/gogo/models.py b/gogo/models.py
index 0e00a67..6767157 100644
--- a/gogo/models.py
+++ b/gogo/models.py
@@ -241,6 +241,24 @@ class BookingRequest(Base, TimestampMixin):
     resolution_note: Mapped[str] = mapped_column(Text, default="")
 
 
+class CallRegistration(Base):
+    """Handoff between the Asterisk dialplan and the voice pipeline (M4).
+
+    The dialplan registers an inbound call over the internal API before ringing
+    anyone; the AudioSocket server looks the UUID up to resolve tenant+caller.
+    """
+
+    __tablename__ = "call_registrations"
+
+    id: Mapped[uuid.UUID] = mapped_column(Uuid, primary_key=True, default=uuid.uuid4)
+    tenant_id: Mapped[uuid.UUID] = mapped_column(ForeignKey("tenants.id"), index=True)
+    caller_msisdn: Mapped[str] = mapped_column(String(30), default="")
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), default=utcnow)
+    answered_by_human: Mapped[bool] = mapped_column(Boolean, default=False)
+    handled_by_agent: Mapped[bool] = mapped_column(Boolean, default=False)
+    finalized: Mapped[bool] = mapped_column(Boolean, default=False)
+
+
 class ActionToken(Base):
     """Signed single-use action link tokens for proposal emails (§9.1)."""
 
diff --git a/gogo/telephony/asterisk.py b/gogo/telephony/asterisk.py
new file mode 100644
index 0000000..79f7480
--- /dev/null
+++ b/gogo/telephony/asterisk.py
@@ -0,0 +1,222 @@
+"""Asterisk configuration generator (M4): pjsip.conf + extensions.conf.
+
+Generated from the DB (tenants, workers) so the dashboard stays the single
+source of truth. Ring-group-first logic (§5.2) lives in the dialplan, but the
+decision (targets + timeout, working hours) comes from the backend via the
+internal API at call time — Asterisk never duplicates tenant config.
+
+Usage:  python -m gogo.telephony.asterisk --out ./asterisk/conf
+Regenerate + `asterisk -rx 'core reload'` whenever workers change.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import sys
+from pathlib import Path
+
+from sqlalchemy import select
+
+from gogo.config import get_settings
+from gogo.db import get_sessionmaker
+from gogo.models import Tenant, Worker
+
+PJSIP_HEADER = """\
+; -- generated by gogo.telephony.asterisk — do not edit by hand --
+
+[global]
+type=global
+user_agent=GogoTelefon
+
+[transport-udp]
+type=transport
+protocol=udp
+bind=0.0.0.0:5060
+"""
+
+ENDPOINT_TEMPLATE = """\
+
+; worker: {name} ({tenant})
+[{username}]
+type=endpoint
+context={context}
+disallow=all
+allow=alaw,ulaw,slin
+auth={username}-auth
+aors={username}
+direct_media=no
+rtp_symmetric=yes
+force_rport=yes
+rewrite_contact=yes
+
+[{username}-auth]
+type=auth
+auth_type=userpass
+username={username}
+password={password}
+
+[{username}]
+type=aor
+max_contacts=3
+remove_existing=yes
+"""
+
+# Test endpoint for M4 software-only testing: a desktop softphone registers as
+# 'test-caller' and dials any number to reach the tenant's inbound flow (§16 M4).
+TEST_CALLER_TEMPLATE = """\
+
+; software-testing caller (softphone) → lands in {context}
+[test-caller-{slug}]
+type=endpoint
+context={context}
+disallow=all
+allow=alaw,ulaw,slin
+auth=test-caller-{slug}-auth
+aors=test-caller-{slug}
+direct_media=no
+rtp_symmetric=yes
+force_rport=yes
+rewrite_contact=yes
+
+[test-caller-{slug}-auth]
+type=auth
+auth_type=userpass
+username=test-caller-{slug}
+password={password}
+
+[test-caller-{slug}]
+type=aor
+max_contacts=3
+remove_existing=yes
+"""
+
+EXTENSIONS_HEADER = """\
+; -- generated by gogo.telefonija.asterisk — do not edit by hand --
+
+[globals]
+GOGO_BACKEND={backend}
+GOGO_VOICE={voice}
+GOGO_TOKEN={token}
+
+[gogo-workers]
+; workers only receive calls in MVP
+exten => _X.,1,Hangup()
+"""
+
+TENANT_CONTEXT_TEMPLATE = """\
+
+[{context}]
+; inbound flow for {tenant_name} (§5.2): register → ring staff → agent fallback
+exten => _[+0-9].,1,NoOp(Gogo inbound ${{EXTEN}} for {slug})
+ same => n,Answer()
+ same => n,Set(REG=${{CURL(${{GOGO_BACKEND}}/internal/calls/register?tenant={slug}&caller=${{URIENCODE(${{CALLERID(num)}})}}&token=${{GOGO_TOKEN}})}})
+ same => n,GotoIf($["${{REG}}"=""]?fallback)
+ same => n,Set(GOGO_UUID=${{CUT(REG,\\,,1)}})
+ same => n,Set(RINGTIME=${{CUT(REG,\\,,2)}})
+ same => n,Set(TARGETS=${{CUT(REG,\\,,3)}})
+ same => n,GotoIf($["${{TARGETS}}"=""]?agent)
+ same => n,Dial(${{TARGETS}},${{RINGTIME}})
+ same => n,GotoIf($["${{DIALSTATUS}}"="ANSWER"]?answered)
+ same => n(agent),AudioSocket(${{GOGO_UUID}},${{GOGO_VOICE}})
+ same => n,Hangup()
+ same => n(answered),Set(RES=${{CURL(${{GOGO_BACKEND}}/internal/calls/${{GOGO_UUID}}/answered?token=${{GOGO_TOKEN}}&duration=${{CDR(billsec)}})}})
+ same => n,Hangup()
+ same => n(fallback),Playback(vm-goodbye)
+ same => n,Hangup()
+
+; caller hung up mid-flow (e.g. while ringing) → missed-call handling (§5.5)
+exten => h,1,GotoIf($["${{GOGO_UUID}}"=""]?done)
+ same => n,Set(RES=${{CURL(${{GOGO_BACKEND}}/internal/calls/${{GOGO_UUID}}/hangup?status=${{DIALSTATUS}}&token=${{GOGO_TOKEN}})}})
+ same => n(done),NoOp(bye)
+"""
+
+
+async def generate(out_dir: Path, backend_url: str = "", voice_addr: str = "") -> list[str]:
+    """Write pjsip.conf + extensions.conf; returns list of written files."""
+    s = get_settings()
+    backend_url = backend_url or "http://backend:8000"
+    voice_addr = voice_addr or f"voice:{s.voice_port}"
+
+    pjsip = [PJSIP_HEADER]
+    extensions = [
+        EXTENSIONS_HEADER.format(backend=backend_url, voice=voice_addr, token=s.internal_token)
+    ]
+
+    async with get_sessionmaker()() as session:
+        tenants = (
+            (
+                await session.execute(
+                    select(Tenant).where(Tenant.status == "active").order_by(Tenant.slug)
+                )
+            )
+            .scalars()
+            .all()
+        )
+        for tenant in tenants:
+            context = f"gogo-inbound-{tenant.slug}"
+            workers = (
+                (
+                    await session.execute(
+                        select(Worker)
+                        .where(Worker.tenant_id == tenant.id, Worker.active.is_(True))
+                        .order_by(Worker.sip_username)
+                    )
+                )
+                .scalars()
+                .all()
+            )
+            for w in workers:
+                pjsip.append(
+                    ENDPOINT_TEMPLATE.format(
+                        name=w.name,
+                        tenant=tenant.slug,
+                        username=w.sip_username,
+                        password=w.sip_password,
+                        context="gogo-workers",
+                    )
+                )
+            pjsip.append(
+                TEST_CALLER_TEMPLATE.format(
+                    slug=tenant.slug,
+                    context=context,
+                    password=f"test-{tenant.widget_public_key[:10]}",
+                )
+            )
+            extensions.append(
+                TENANT_CONTEXT_TEMPLATE.format(
+                    context=context, tenant_name=tenant.name, slug=tenant.slug
+                )
+            )
+
+    out_dir.mkdir(parents=True, exist_ok=True)
+    written = []
+    for filename, content in [
+        ("pjsip.conf", "".join(pjsip)),
+        ("extensions.conf", "".join(extensions)),
+    ]:
+        path = out_dir / filename
+        path.write_text(content)
+        written.append(str(path))
+    return written
+
+
+def main() -> None:
+    out = Path("./asterisk/conf")
+    backend = ""
+    voice = ""
+    args = sys.argv[1:]
+    while args:
+        arg = args.pop(0)
+        if arg == "--out" and args:
+            out = Path(args.pop(0))
+        elif arg == "--backend" and args:
+            backend = args.pop(0)
+        elif arg == "--voice" and args:
+            voice = args.pop(0)
+    written = asyncio.run(generate(out, backend, voice))
+    for path in written:
+        print(f"wrote {path}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/gogo/voice/__init__.py b/gogo/voice/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/gogo/voice/audio.py b/gogo/voice/audio.py
new file mode 100644
index 0000000..7f99f69
--- /dev/null
+++ b/gogo/voice/audio.py
@@ -0,0 +1,94 @@
+"""Audio helpers for the voice pipeline.
+
+Wire format everywhere: SLIN — signed 16-bit LE mono PCM @ 8 kHz (AudioSocket).
+Whisper wants 16 kHz; TTS providers are asked for 8 kHz where possible and
+resampled otherwise.
+"""
+
+from __future__ import annotations
+
+import audioop
+import io
+import math
+import wave
+
+SAMPLE_RATE = 8000
+SAMPLE_WIDTH = 2  # bytes, 16-bit
+FRAME_MS = 20
+FRAME_BYTES = SAMPLE_RATE * SAMPLE_WIDTH * FRAME_MS // 1000  # 320
+
+
+def resample(pcm: bytes, from_rate: int, to_rate: int) -> bytes:
+    if from_rate == to_rate or not pcm:
+        return pcm
+    out, _ = audioop.ratecv(pcm, SAMPLE_WIDTH, 1, from_rate, to_rate, None)
+    return out
+
+
+def rms(pcm: bytes) -> int:
+    return audioop.rms(pcm, SAMPLE_WIDTH) if pcm else 0
+
+
+def duration_s(pcm: bytes, rate: int = SAMPLE_RATE) -> float:
+    return len(pcm) / (rate * SAMPLE_WIDTH)
+
+
+def silence(ms: int, rate: int = SAMPLE_RATE) -> bytes:
+    return b"\x00" * (rate * SAMPLE_WIDTH * ms // 1000)
+
+
+def tone(freq: int, ms: int, rate: int = SAMPLE_RATE, amplitude: int = 12000) -> bytes:
+    """Test helper: sine tone (registers as speech for the energy VAD)."""
+    n = rate * ms // 1000
+    return b"".join(
+        int(amplitude * math.sin(2 * math.pi * freq * i / rate)).to_bytes(
+            2, "little", signed=True
+        )
+        for i in range(n)
+    )
+
+
+def pcm_to_wav(pcm: bytes, rate: int = SAMPLE_RATE) -> bytes:
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as w:
+        w.setnchannels(1)
+        w.setsampwidth(SAMPLE_WIDTH)
+        w.setframerate(rate)
+        w.writeframes(pcm)
+    return buf.getvalue()
+
+
+def mix(a: bytes, b: bytes) -> bytes:
+    """Mix two PCM streams (unequal lengths ok)."""
+    if len(a) < len(b):
+        a, b = b, a
+    if not b:
+        return a
+    mixed = audioop.add(a[: len(b)], b, SAMPLE_WIDTH)
+    return mixed + a[len(b):]
+
+
+class CallRecorder:
+    """Accumulates caller + agent audio on a shared timeline and writes a WAV."""
+
+    def __init__(self) -> None:
+        self._caller = bytearray()
+        self._agent = bytearray()
+
+    def add_caller(self, pcm: bytes) -> None:
+        self._caller.extend(pcm)
+        # keep agent track in sync (silence while caller talks)
+        if len(self._agent) < len(self._caller):
+            self._agent.extend(b"\x00" * (len(self._caller) - len(self._agent)))
+
+    def add_agent(self, pcm: bytes) -> None:
+        self._agent.extend(pcm)
+        if len(self._caller) < len(self._agent):
+            self._caller.extend(b"\x00" * (len(self._agent) - len(self._caller)))
+
+    def to_wav(self) -> bytes:
+        return pcm_to_wav(mix(bytes(self._caller), bytes(self._agent)))
+
+    @property
+    def seconds(self) -> float:
+        return duration_s(bytes(self._caller))
diff --git a/gogo/voice/call.py b/gogo/voice/call.py
new file mode 100644
index 0000000..82a494e
--- /dev/null
+++ b/gogo/voice/call.py
@@ -0,0 +1,246 @@
+"""One agent-handled phone call: audio loop, barge-in, recording, metering (M4).
+
+The transport hands us paced 20 ms SLIN frames from Asterisk (AudioSocket) and
+accepts frames to play back. Flow per §6.2: greeting → collect utterance (VAD)
+→ STT → agent turn (M2 loop, same tools as chat) → TTS → play, with barge-in
+(caller speech stops playback). Unintelligible audio reaches the agent as
+'[nerazumljivo]' so the two-attempts rule (§6.4/A.6) lives in the prompt.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+import time
+import uuid
+from dataclasses import dataclass
+
+from sqlalchemy import select
+
+from gogo.agent.loop import AgentConversation
+from gogo.config import get_settings
+from gogo.db import get_sessionmaker
+from gogo.metering import record_agent_call
+from gogo.models import Call, CallRegistration, Tenant, utcnow
+from gogo.voice.audio import FRAME_BYTES, FRAME_MS, CallRecorder
+from gogo.voice.stt import get_stt
+from gogo.voice.tts import get_tts
+from gogo.voice.vad import UtteranceCollector
+
+log = logging.getLogger("gogo.voice.call")
+
+UNCLEAR_MARKER = "[nerazumljivo]"
+GOODBYE_TIMEOUT = (
+    "Izgleda da veza ne radi. Salon će vidjeti vaš broj i javiti vam se. Prijatno!"
+)
+
+BARGE_IN_FRAMES = 6  # 120 ms of speech during playback stops it
+
+
+@dataclass
+class CallLimits:
+    inactivity_s: float = 30.0
+    max_call_s: float = 600.0
+    max_turns: int = 30
+
+
+class Transport:
+    """What CallSession needs from the wire (implemented by AudioSocket + fakes)."""
+
+    async def read_frame(self) -> bytes | None:
+        """Next inbound 20 ms frame, or None on hangup."""
+        raise NotImplementedError
+
+    async def send_audio(self, pcm: bytes) -> None:
+        raise NotImplementedError
+
+    async def hangup(self) -> None:
+        raise NotImplementedError
+
+
+class CallSession:
+    def __init__(
+        self,
+        transport: Transport,
+        registration_id: uuid.UUID,
+        *,
+        limits: CallLimits | None = None,
+        pace: float = FRAME_MS / 1000,
+    ):
+        self.transport = transport
+        self.registration_id = registration_id
+        self.limits = limits or CallLimits()
+        self.pace = pace
+        self.recorder = CallRecorder()
+        self.collector = UtteranceCollector()
+        self._started = time.monotonic()
+
+    async def run(self) -> None:
+        async with get_sessionmaker()() as session:
+            reg = (
+                await session.execute(
+                    select(CallRegistration).where(CallRegistration.id == self.registration_id)
+                )
+            ).scalar_one_or_none()
+            if reg is None:
+                log.error("no registration for call %s — hanging up", self.registration_id)
+                await self.transport.hangup()
+                return
+            tenant = (
+                await session.execute(select(Tenant).where(Tenant.id == reg.tenant_id))
+            ).scalar_one()
+            reg.handled_by_agent = True
+
+            call = Call(
+                tenant_id=tenant.id,
+                caller_msisdn=reg.caller_msisdn,
+                started_at=utcnow(),
+            )
+            session.add(call)
+            await session.flush()
+
+            convo = AgentConversation(
+                session,
+                tenant,
+                channel="voice",
+                caller_phone=reg.caller_msisdn,
+                call_id=call.id,
+            )
+            tts, voice = get_tts(tenant.tts_voice)
+            stt = get_stt()
+
+            try:
+                greeting = await convo.greeting()
+                await session.commit()
+                await self._speak(tts, voice, greeting)
+
+                for _turn in range(self.limits.max_turns):
+                    if time.monotonic() - self._started > self.limits.max_call_s:
+                        log.warning("max call duration hit (tenant %s)", tenant.slug)
+                        break
+                    utterance = await self._collect_utterance()
+                    if utterance is None:  # hangup
+                        break
+                    if utterance == b"":  # inactivity
+                        await self._speak(tts, voice, GOODBYE_TIMEOUT)
+                        break
+
+                    t0 = time.monotonic()
+                    text = await stt.transcribe(utterance)
+                    stt_ms = int((time.monotonic() - t0) * 1000)
+
+                    t0 = time.monotonic()
+                    reply = await convo.user_turn(text or UNCLEAR_MARKER)
+                    llm_ms = int((time.monotonic() - t0) * 1000)
+                    await session.commit()
+
+                    t0 = time.monotonic()
+                    interrupted = await self._speak(tts, voice, reply)
+                    tts_ms = int((time.monotonic() - t0) * 1000)
+
+                    trace = call.trace or {}
+                    turns = trace.setdefault("turns", [])
+                    turns.append(
+                        {"stt_ms": stt_ms, "llm_ms": llm_ms, "speak_ms": tts_ms,
+                         "interrupted": interrupted, "chars": len(reply)}
+                    )
+                    call.trace = dict(trace)
+
+                    if self._agent_closed(convo):
+                        break
+            except Exception:  # noqa: BLE001 — a crash must still finalize the call
+                log.exception("call session error (tenant %s)", tenant.slug)
+            finally:
+                await self._finalize(session, tenant, call, convo, reg)
+                await self.transport.hangup()
+
+    # -- audio helpers ------------------------------------------------------
+
+    async def _speak(self, tts, voice: str, text: str) -> bool:
+        """Synthesize and play; returns True if the caller barged in."""
+        if not text:
+            return False
+        try:
+            pcm = await tts.synthesize(text, voice)
+        except Exception:  # noqa: BLE001
+            log.exception("TTS failed — continuing without audio")
+            return False
+
+        voiced_streak = 0
+        for i in range(0, len(pcm), FRAME_BYTES):
+            frame = pcm[i : i + FRAME_BYTES]
+            await self.transport.send_audio(frame)
+            self.recorder.add_agent(frame)
+            if self.pace:
+                await asyncio.sleep(self.pace)
+            # barge-in: watch inbound audio while we play
+            inbound = await self._poll_frame()
+            if inbound is not None:
+                self.recorder.add_caller(inbound)
+                if self.collector.is_speech(inbound):
+                    voiced_streak += 1
+                    if voiced_streak >= BARGE_IN_FRAMES:
+                        self.collector.reset()
+                        self.collector.feed(inbound)
+                        return True
+                else:
+                    voiced_streak = 0
+        return False
+
+    async def _poll_frame(self) -> bytes | None:
+        try:
+            return await asyncio.wait_for(self.transport.read_frame(), timeout=0.001)
+        except TimeoutError:
+            return None
+
+    async def _collect_utterance(self) -> bytes | None:
+        """None = hangup, b'' = inactivity timeout, else utterance PCM."""
+        deadline = time.monotonic() + self.limits.inactivity_s
+        while True:
+            timeout = deadline - time.monotonic()
+            if timeout <= 0:
+                return b""
+            try:
+                frame = await asyncio.wait_for(self.transport.read_frame(), timeout=timeout)
+            except TimeoutError:
+                return b""
+            if frame is None:
+                return None
+            self.recorder.add_caller(frame)
+            utterance = self.collector.feed(frame)
+            if utterance is not None:
+                return utterance
+
+    def _agent_closed(self, convo: AgentConversation) -> bool:
+        """The closing line ends the call from our side (short calls §2.5)."""
+        last = next((t.text for t in reversed(convo.turns) if t.role == "assistant"), "")
+        lowered = last.lower()
+        return ("prijatno" in lowered or "doviđenja" in lowered) and bool(
+            convo.executor.created_request_id or convo.executor.took_message
+        )
+
+    async def _finalize(self, session, tenant, call, convo, reg) -> None:
+        try:
+            call.duration_s = int(time.monotonic() - self._started)
+            call.agent_seconds_charged = call.duration_s
+            call.outcome = convo.outcome
+            call.transcript = convo.transcript
+
+            rec_dir = get_settings().recordings_dir
+            os.makedirs(rec_dir, exist_ok=True)
+            path = os.path.join(rec_dir, f"{call.id}.wav")
+            with open(path, "wb") as f:
+                f.write(self.recorder.to_wav())
+            call.recording_path = path
+
+            reg.finalized = True
+            await record_agent_call(session, tenant, call.agent_seconds_charged)
+            await session.commit()
+            log.info(
+                "call %s finalized: %ss outcome=%s tenant=%s",
+                call.id, call.duration_s, call.outcome, tenant.slug,
+            )
+        except Exception:  # noqa: BLE001
+            log.exception("failed to finalize call %s", call.id)
+            await session.rollback()
diff --git a/gogo/voice/server.py b/gogo/voice/server.py
new file mode 100644
index 0000000..a31f54f
--- /dev/null
+++ b/gogo/voice/server.py
@@ -0,0 +1,113 @@
+"""AudioSocket server — Asterisk connects here per agent-handled call (M4).
+
+AudioSocket wire protocol (Asterisk `AudioSocket()` app):
+    1 byte kind | 2 bytes big-endian payload length | payload
+    kind 0x00 = terminate, 0x01 = UUID (16 raw bytes), 0x10 = SLIN audio,
+    0xff = error. Audio is 16-bit LE mono PCM @ 8 kHz, ~20 ms per message.
+
+Run:  python -m gogo.voice.server
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import uuid
+
+from gogo.config import get_settings
+from gogo.voice.call import CallSession, Transport
+
+log = logging.getLogger("gogo.voice.server")
+
+KIND_TERMINATE = 0x00
+KIND_UUID = 0x01
+KIND_AUDIO = 0x10
+KIND_ERROR = 0xFF
+
+
+class AudioSocketTransport(Transport):
+    def __init__(self, reader: asyncio.StreamReader, writer: asyncio.StreamWriter):
+        self.reader = reader
+        self.writer = writer
+        self._closed = False
+
+    async def read_message(self) -> tuple[int, bytes] | None:
+        try:
+            header = await self.reader.readexactly(3)
+        except (asyncio.IncompleteReadError, ConnectionResetError):
+            return None
+        kind = header[0]
+        length = int.from_bytes(header[1:3], "big")
+        payload = b""
+        if length:
+            try:
+                payload = await self.reader.readexactly(length)
+            except (asyncio.IncompleteReadError, ConnectionResetError):
+                return None
+        return kind, payload
+
+    async def read_frame(self) -> bytes | None:
+        """Next audio frame; skips non-audio messages; None on hangup/close."""
+        while True:
+            msg = await self.read_message()
+            if msg is None:
+                return None
+            kind, payload = msg
+            if kind == KIND_AUDIO:
+                return payload
+            if kind == KIND_TERMINATE:
+                return None
+            if kind == KIND_ERROR:
+                log.warning("audiosocket error frame: %s", payload[:64])
+
+    async def send_audio(self, pcm: bytes) -> None:
+        if self._closed:
+            return
+        try:
+            self.writer.write(bytes([KIND_AUDIO]) + len(pcm).to_bytes(2, "big") + pcm)
+            await self.writer.drain()
+        except (ConnectionResetError, BrokenPipeError):
+            self._closed = True
+
+    async def hangup(self) -> None:
+        if self._closed:
+            return
+        self._closed = True
+        try:
+            self.writer.write(bytes([KIND_TERMINATE, 0, 0]))
+            await self.writer.drain()
+            self.writer.close()
+        except (ConnectionResetError, BrokenPipeError):
+            pass
+
+
+async def handle_connection(reader: asyncio.StreamReader, writer: asyncio.StreamWriter) -> None:
+    transport = AudioSocketTransport(reader, writer)
+    msg = await transport.read_message()
+    if msg is None or msg[0] != KIND_UUID or len(msg[1]) != 16:
+        log.warning("connection without UUID handshake — dropping")
+        await transport.hangup()
+        return
+    registration_id = uuid.UUID(bytes=msg[1])
+    log.info("call connected: %s", registration_id)
+    session = CallSession(transport, registration_id)
+    await session.run()
+
+
+async def serve() -> None:
+    s = get_settings()
+    server = await asyncio.start_server(handle_connection, s.voice_host, s.voice_port)
+    addrs = ", ".join(str(sock.getsockname()) for sock in server.sockets)
+    log.info("AudioSocket server listening on %s", addrs)
+    # health heartbeat file for the admin panel / monitoring (M6)
+    async with server:
+        await server.serve_forever()
+
+
+def main() -> None:
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s %(message)s")
+    asyncio.run(serve())
+
+
+if __name__ == "__main__":
+    main()
diff --git a/gogo/voice/stt.py b/gogo/voice/stt.py
new file mode 100644
index 0000000..187cd43
--- /dev/null
+++ b/gogo/voice/stt.py
@@ -0,0 +1,81 @@
+"""STTProvider interface — faster-whisper implementation + test fake (§4).
+
+Language hint 'sr' covers spoken Bosnian/Serbian/Croatian (§6.1). The model
+runs on GPU in the founder's DC (large-v3); CPU with a small model for dev.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+from typing import Protocol
+
+from gogo.voice.audio import SAMPLE_RATE, resample
+
+log = logging.getLogger("gogo.voice.stt")
+
+
+class STTProvider(Protocol):
+    async def transcribe(self, pcm_8k: bytes) -> str:
+        """SLIN 8 kHz PCM → text (empty string when nothing intelligible)."""
+        ...
+
+
+class WhisperSTT:
+    """faster-whisper; model size/device via env (GOGO_WHISPER_MODEL, GOGO_WHISPER_DEVICE)."""
+
+    def __init__(self, model_name: str | None = None, device: str | None = None):
+        from faster_whisper import WhisperModel  # heavy import, voice extra
+
+        model_name = model_name or os.environ.get("GOGO_WHISPER_MODEL", "small")
+        device = device or os.environ.get("GOGO_WHISPER_DEVICE", "auto")
+        compute = "float16" if device == "cuda" else "int8"
+        log.info("loading whisper model=%s device=%s", model_name, device)
+        self._model = WhisperModel(model_name, device=device, compute_type=compute)
+        self._lock = asyncio.Lock()
+
+    async def transcribe(self, pcm_8k: bytes) -> str:
+        import numpy as np
+
+        pcm_16k = resample(pcm_8k, SAMPLE_RATE, 16000)
+        audio = np.frombuffer(pcm_16k, dtype=np.int16).astype(np.float32) / 32768.0
+
+        def run() -> str:
+            segments, _info = self._model.transcribe(
+                audio,
+                language="sr",  # covers bs/sr/hr as one spoken language (§6.1)
+                beam_size=5,
+                vad_filter=True,
+            )
+            return " ".join(s.text.strip() for s in segments).strip()
+
+        async with self._lock:  # one decode at a time per worker process
+            return await asyncio.get_running_loop().run_in_executor(None, run)
+
+
+class FakeSTT:
+    """Test fake: returns queued texts, one per transcribe() call."""
+
+    def __init__(self, texts: list[str] | None = None):
+        self.texts = list(texts or [])
+        self.received: list[bytes] = []
+
+    async def transcribe(self, pcm_8k: bytes) -> str:
+        self.received.append(pcm_8k)
+        return self.texts.pop(0) if self.texts else ""
+
+
+_stt: STTProvider | None = None
+
+
+def get_stt() -> STTProvider:
+    global _stt
+    if _stt is None:
+        _stt = WhisperSTT()
+    return _stt
+
+
+def set_stt(provider: STTProvider | None) -> None:
+    global _stt
+    _stt = provider
diff --git a/gogo/voice/tts.py b/gogo/voice/tts.py
new file mode 100644
index 0000000..a412045
--- /dev/null
+++ b/gogo/voice/tts.py
@@ -0,0 +1,107 @@
+"""TTSProvider interface (§4): Azure Neural + ElevenLabs Flash + test fake.
+
+Both return SLIN 8 kHz PCM ready for AudioSocket. Default provider/voice comes
+from settings; per-tenant voice ids look like "azure:sr-RS-SophieNeural" or
+"elevenlabs:<voice_id>".
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Protocol
+from xml.sax.saxutils import escape
+
+import httpx
+
+from gogo.config import get_settings
+from gogo.voice.audio import SAMPLE_RATE, resample, tone
+
+log = logging.getLogger("gogo.voice.tts")
+
+DEFAULT_AZURE_VOICE = "sr-RS-SophieNeural"
+
+
+class TTSProvider(Protocol):
+    async def synthesize(self, text: str, voice: str = "") -> bytes:
+        """Text → SLIN 8 kHz PCM."""
+        ...
+
+
+class AzureTTS:
+    async def synthesize(self, text: str, voice: str = "") -> bytes:
+        s = get_settings()
+        voice = voice or DEFAULT_AZURE_VOICE
+        ssml = (
+            f"<speak version='1.0' xml:lang='sr-RS'>"
+            f"<voice name='{voice}'>{escape(text)}</voice></speak>"
+        )
+        url = (
+            f"https://{s.azure_speech_region}.tts.speech.microsoft.com/"
+            "cognitiveservices/v1"
+        )
+        async with httpx.AsyncClient(timeout=15) as client:
+            resp = await client.post(
+                url,
+                content=ssml.encode(),
+                headers={
+                    "Ocp-Apim-Subscription-Key": s.azure_speech_key,
+                    "Content-Type": "application/ssml+xml",
+                    "X-Microsoft-OutputFormat": "raw-8khz-16bit-mono-pcm",
+                    "User-Agent": "gogotelefon",
+                },
+            )
+            resp.raise_for_status()
+            return resp.content
+
+
+class ElevenLabsTTS:
+    async def synthesize(self, text: str, voice: str = "") -> bytes:
+        s = get_settings()
+        voice_id = voice or "JBFqnCBsd6RMkjVDRZzb"
+        async with httpx.AsyncClient(timeout=20) as client:
+            resp = await client.post(
+                f"https://api.elevenlabs.io/v1/text-to-speech/{voice_id}",
+                params={"output_format": "pcm_16000"},
+                headers={"xi-api-key": s.elevenlabs_api_key},
+                json={"text": text, "model_id": "eleven_flash_v2_5"},
+            )
+            resp.raise_for_status()
+            return resample(resp.content, 16000, SAMPLE_RATE)
+
+
+class FakeTTS:
+    """Test fake: N ms of tone per character (deterministic, VAD-audible)."""
+
+    def __init__(self, ms_per_char: int = 5):
+        self.ms_per_char = ms_per_char
+        self.spoken: list[str] = []
+
+    async def synthesize(self, text: str, voice: str = "") -> bytes:
+        self.spoken.append(text)
+        return tone(440, max(40, self.ms_per_char * len(text)))
+
+
+def make_tts(provider_and_voice: str = "") -> tuple[TTSProvider, str]:
+    """Resolve a tenant voice id 'provider:voice' → (provider instance, voice)."""
+    s = get_settings()
+    provider_name, _, voice = (provider_and_voice or "").partition(":")
+    if not voice:
+        provider_name, voice = s.tts_provider, ""
+    if provider_name == "elevenlabs":
+        return ElevenLabsTTS(), voice
+    return AzureTTS(), voice or DEFAULT_AZURE_VOICE
+
+
+_tts: TTSProvider | None = None
+
+
+def get_tts(tenant_voice: str = "") -> tuple[TTSProvider, str]:
+    if _tts is not None:  # test override
+        _, _, voice = (tenant_voice or "").partition(":")
+        return _tts, voice
+    return make_tts(tenant_voice)
+
+
+def set_tts(provider: TTSProvider | None) -> None:
+    global _tts
+    _tts = provider
diff --git a/gogo/voice/vad.py b/gogo/voice/vad.py
new file mode 100644
index 0000000..946442e
--- /dev/null
+++ b/gogo/voice/vad.py
@@ -0,0 +1,83 @@
+"""Utterance detection: adaptive energy VAD with hangover.
+
+Deliberately simple and dependency-free; swap for silero/webrtcvad later if the
+Phase-0 PoC shows it is needed. Operates on 20 ms SLIN frames.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from gogo.voice.audio import FRAME_MS, rms
+
+
+@dataclass
+class VadConfig:
+    # energy threshold = max(min_threshold, noise_floor * ratio)
+    min_threshold: int = 500
+    noise_ratio: float = 3.0
+    start_ms: int = 120  # this much voiced audio starts an utterance
+    end_ms: int = 700  # this much silence ends it
+    max_utterance_ms: int = 15000  # hard cap
+    pre_roll_ms: int = 200  # audio kept from before the trigger
+
+
+@dataclass
+class UtteranceCollector:
+    """Feed 20 ms frames; returns the full utterance PCM when one completes."""
+
+    config: VadConfig = field(default_factory=VadConfig)
+    _noise_floor: float = 200.0
+    _voiced_ms: int = 0
+    _silence_ms: int = 0
+    _in_utterance: bool = False
+    _buffer: bytearray = field(default_factory=bytearray)
+    _pre_roll: bytearray = field(default_factory=bytearray)
+
+    def is_speech(self, frame: bytes) -> bool:
+        energy = rms(frame)
+        threshold = max(self.config.min_threshold, self._noise_floor * self.config.noise_ratio)
+        speech = energy > threshold
+        if not speech:
+            # slowly track the noise floor on non-speech frames
+            self._noise_floor = 0.95 * self._noise_floor + 0.05 * energy
+        return speech
+
+    def feed(self, frame: bytes) -> bytes | None:
+        """Returns utterance PCM when a complete utterance is detected, else None."""
+        speech = self.is_speech(frame)
+
+        if not self._in_utterance:
+            self._pre_roll.extend(frame)
+            max_pre = self.config.pre_roll_ms * len(frame) // FRAME_MS
+            if len(self._pre_roll) > max_pre:
+                del self._pre_roll[: len(self._pre_roll) - max_pre]
+            if speech:
+                self._voiced_ms += FRAME_MS
+                if self._voiced_ms >= self.config.start_ms:
+                    self._in_utterance = True
+                    self._buffer = bytearray(self._pre_roll)
+                    self._silence_ms = 0
+            else:
+                self._voiced_ms = 0
+            return None
+
+        self._buffer.extend(frame)
+        if speech:
+            self._silence_ms = 0
+        else:
+            self._silence_ms += FRAME_MS
+
+        utterance_ms = len(self._buffer) * FRAME_MS // len(frame)
+        if self._silence_ms >= self.config.end_ms or utterance_ms >= self.config.max_utterance_ms:
+            utterance = bytes(self._buffer)
+            self.reset()
+            return utterance
+        return None
+
+    def reset(self) -> None:
+        self._in_utterance = False
+        self._voiced_ms = 0
+        self._silence_ms = 0
+        self._buffer = bytearray()
+        self._pre_roll = bytearray()
diff --git a/tests/test_voice.py b/tests/test_voice.py
new file mode 100644
index 0000000..5348628
--- /dev/null
+++ b/tests/test_voice.py
@@ -0,0 +1,512 @@
+"""M4 voice pipeline: VAD, audio utils, full fake-call e2e (greeting → utterance
+→ STT → agent → TTS → booking request → finalize with recording + metering),
+barge-in, internal dialplan API, Asterisk config generation."""
+
+import asyncio
+import uuid
+from datetime import datetime, timedelta
+from zoneinfo import ZoneInfo
+
+import pytest
+from sqlalchemy import select
+
+from gogo.agent.llm import LLMResponse, ScriptedLLM, ToolUse, set_llm
+from gogo.models import BookingRequest, Call, CallRegistration, UsageCounter, Worker
+from gogo.voice.audio import FRAME_BYTES, CallRecorder, mix, pcm_to_wav, silence, tone
+from gogo.voice.call import CallLimits, CallSession, Transport
+from gogo.voice.stt import FakeSTT, set_stt
+from gogo.voice.tts import FakeTTS, set_tts
+from gogo.voice.vad import UtteranceCollector
+
+TZ = ZoneInfo("Europe/Sarajevo")
+
+
+def frames(pcm: bytes) -> list[bytes]:
+    return [pcm[i : i + FRAME_BYTES] for i in range(0, len(pcm), FRAME_BYTES)]
+
+
+# -- audio + vad units ----------------------------------------------------------
+
+
+def test_vad_detects_utterance_between_silence():
+    collector = UtteranceCollector()
+    utterance = None
+    for frame in frames(silence(500) + tone(440, 1000) + silence(800)):
+        result = collector.feed(frame)
+        if result is not None:
+            utterance = result
+    assert utterance is not None
+    assert len(utterance) >= len(tone(440, 900))  # roughly the voiced part (+pre-roll)
+
+
+def test_vad_ignores_pure_silence():
+    collector = UtteranceCollector()
+    assert all(collector.feed(f) is None for f in frames(silence(3000)))
+
+
+def test_recorder_produces_wav():
+    rec = CallRecorder()
+    rec.add_agent(tone(440, 100))
+    rec.add_caller(tone(220, 100))
+    wav = rec.to_wav()
+    assert wav[:4] == b"RIFF"
+    assert rec.seconds > 0
+
+
+def test_mix_unequal_lengths():
+    a, b = tone(440, 100), tone(220, 40)
+    assert len(mix(a, b)) == len(a)
+    assert pcm_to_wav(mix(a, b))[:4] == b"RIFF"
+
+
+# -- fake transport ---------------------------------------------------------------
+
+
+class FakeCallerTransport(Transport):
+    """Scripted caller: plays queued PCM as 20ms frames, then silence forever."""
+
+    def __init__(self, pcm_script: bytes, hang_up_after_script: bool = False):
+        self.frames = frames(pcm_script)
+        self.hang_up_after_script = hang_up_after_script
+        self.sent_audio = bytearray()
+        self.hung_up = False
+
+    async def read_frame(self) -> bytes | None:
+        await asyncio.sleep(0)
+        if self.hung_up:
+            return None
+        if self.frames:
+            return self.frames.pop(0)
+        if self.hang_up_after_script:
+            return None
+        return silence(20)
+
+    async def send_audio(self, pcm: bytes) -> None:
+        self.sent_audio.extend(pcm)
+
+    async def hangup(self) -> None:
+        self.hung_up = True
+
+
+@pytest.fixture
+def voice_fakes(tmp_path, monkeypatch):
+    monkeypatch.setenv("GOGO_RECORDINGS_DIR", str(tmp_path / "rec"))
+    from gogo.config import get_settings
+
+    get_settings.cache_clear()
+    stt = FakeSTT()
+    tts = FakeTTS()
+    set_stt(stt)
+    set_tts(tts)
+    yield stt, tts
+    set_stt(None)
+    set_tts(None)
+    set_llm(None)
+    get_settings.cache_clear()
+
+
+async def make_registration(session, tenant, caller="+38765123456") -> CallRegistration:
+    reg = CallRegistration(tenant_id=tenant.id, caller_msisdn=caller)
+    session.add(reg)
+    await session.commit()
+    _ = reg.id  # eagerly load before any expire_all in tests
+    return reg
+
+
+UTTERANCE = silence(200) + tone(440, 900) + silence(800)
+
+
+# -- e2e ---------------------------------------------------------------------------
+
+
+async def test_full_call_creates_booking_and_finalizes(
+    session, tenant, emails, sms, voice_fakes, clean_mock_provider
+):
+    stt, tts = voice_fakes
+    from gogo.models import Service
+
+    svc = (
+        await session.execute(select(Service).where(Service.tenant_id == tenant.id))
+    ).scalars().first()
+    svc_id = str(svc.id)
+
+    start = datetime.now(TZ).replace(hour=17, minute=0, second=0, microsecond=0) + timedelta(days=2)
+    stt.texts = ["Htjela bih zakazati šišanje, Amra Hodžić, broj s kojeg zovem."]
+    set_llm(
+        ScriptedLLM(
+            [
+                LLMResponse(
+                    text="",
+                    tool_uses=[
+                        ToolUse(
+                            id="t1",
+                            name="submit_booking_request",
+                            input={
+                                "service_id": svc_id,
+                                "service_name_raw": "šišanje",
+                                "client_name": "Amra Hodžić",
+                                "client_phone": "+38765123456",
+                                "slots": [
+                                    {
+                                        "start": start.isoformat(),
+                                        "end": (start + timedelta(minutes=45)).isoformat(),
+                                    }
+                                ],
+                                "summary": "Šišanje.",
+                            },
+                        )
+                    ],
+                    stop_reason="tool_use",
+                ),
+                LLMResponse(
+                    text="Prosljeđujem zahtjev salonu — kontaktiraće vas. Hvala i prijatno!"
+                ),
+            ]
+        )
+    )
+
+    reg = await make_registration(session, tenant)
+    transport = FakeCallerTransport(UTTERANCE)
+    call_session = CallSession(
+        transport, reg.id, pace=0, limits=CallLimits(inactivity_s=5, max_call_s=60)
+    )
+    await call_session.run()
+
+    # greeting + reply spoken via TTS
+    assert len(tts.spoken) == 2
+    assert "razgovor se snima" in tts.spoken[0]
+    assert "prijatno" in tts.spoken[1].lower()
+    assert len(transport.sent_audio) > 0
+    assert transport.hung_up  # agent closed the call after the closing line
+
+    # STT got the utterance audio
+    assert len(stt.received) == 1
+    assert len(stt.received[0]) > FRAME_BYTES * 10
+
+    # booking request created + proposal email
+    session.expire_all()
+    req = (await session.execute(select(BookingRequest))).scalar_one()
+    assert req.client_name == "Amra Hodžić"
+    assert req.source == "voice"
+    assert len(emails) == 1
+
+    # call finalized: transcript, recording, outcome, metering
+    call = (await session.execute(select(Call))).scalar_one()
+    assert call.outcome == "request_created"
+    assert call.transcript[0]["role"] == "assistant"
+    assert any(t["role"] == "user" for t in call.transcript)
+    assert call.recording_path and call.recording_path.endswith(".wav")
+    import os
+
+    assert os.path.exists(call.recording_path)
+    assert call.trace["turns"][0]["stt_ms"] >= 0
+
+    counter = (await session.execute(select(UsageCounter))).scalar_one()
+    assert counter.calls == 1
+
+    reg2 = (
+        await session.execute(
+            select(CallRegistration).where(CallRegistration.id == call_session.registration_id)
+        )
+    ).scalar_one()
+    assert reg2.handled_by_agent and reg2.finalized
+
+
+async def test_unclear_speech_reaches_agent_as_marker(
+    session, tenant, emails, sms, voice_fakes
+):
+    stt, tts = voice_fakes
+    stt.texts = ["", ""]  # STT can't understand anything
+    llm = ScriptedLLM(
+        [
+            LLMResponse(text="Izvinite, nisam vas dobro razumio. Možete li ponoviti?"),
+            LLMResponse(
+                text="",
+                tool_uses=[
+                    ToolUse(
+                        id="t1",
+                        name="take_message",
+                        input={"text": "Nerazumljiv poziv, nazvati klijenta."},
+                    )
+                ],
+                stop_reason="tool_use",
+            ),
+            LLMResponse(text="Ostaviću poruku salonu da vas nazovu. Prijatno!"),
+        ]
+    )
+    set_llm(llm)
+
+    reg = await make_registration(session, tenant, caller="+38765999888")
+    transport = FakeCallerTransport(UTTERANCE + UTTERANCE)
+    call_session = CallSession(
+        transport, reg.id, pace=0, limits=CallLimits(inactivity_s=5, max_call_s=60)
+    )
+    await call_session.run()
+
+    # both user turns arrived as the unclear marker
+    user_msgs = [m for m in llm.calls[-1]["messages"] if m["role"] == "user" and isinstance(m["content"], str)]
+    assert all(m["content"] == "[nerazumljivo]" for m in user_msgs)
+
+    session.expire_all()
+    call = (await session.execute(select(Call))).scalar_one()
+    assert call.outcome == "message_taken"
+
+
+async def test_caller_hangup_mid_call_finalizes(session, tenant, emails, sms, voice_fakes):
+    stt, tts = voice_fakes
+    set_llm(ScriptedLLM([]))
+    reg = await make_registration(session, tenant)
+    transport = FakeCallerTransport(silence(200), hang_up_after_script=True)
+    call_session = CallSession(transport, reg.id, pace=0, limits=CallLimits(inactivity_s=5))
+    await call_session.run()
+
+    session.expire_all()
+    call = (await session.execute(select(Call))).scalar_one()
+    assert call.outcome == "abandoned"  # greeting only, no user turn
+    assert call.recording_path
+
+
+async def test_inactivity_timeout_says_goodbye(session, tenant, emails, sms, voice_fakes):
+    stt, tts = voice_fakes
+    set_llm(ScriptedLLM([]))
+    reg = await make_registration(session, tenant)
+    transport = FakeCallerTransport(b"")  # silent caller
+    call_session = CallSession(
+        transport, reg.id, pace=0, limits=CallLimits(inactivity_s=0.2)
+    )
+    await call_session.run()
+    assert any("veza ne radi" in t for t in tts.spoken)
+    assert transport.hung_up
+
+
+# -- barge-in ---------------------------------------------------------------------
+
+
+class BargeInTransport(FakeCallerTransport):
+    """Starts speaking as soon as the agent starts playing audio."""
+
+    def __init__(self):
+        super().__init__(b"")
+        self.started_speaking = False
+
+    async def send_audio(self, pcm: bytes) -> None:
+        await super().send_audio(pcm)
+        if not self.started_speaking:
+            self.started_speaking = True
+            self.frames = frames(tone(440, 1500) + silence(800))
+
+
+async def test_barge_in_interrupts_playback(session, tenant, emails, sms, voice_fakes):
+    stt, tts = voice_fakes
+    tts.ms_per_char = 40  # long greeting so there is something to interrupt
+    stt.texts = ["Halo, samo pitanje."]
+    set_llm(ScriptedLLM([LLMResponse(text="Izvolite, recite.")]))
+
+    reg = await make_registration(session, tenant)
+    transport = BargeInTransport()
+    call_session = CallSession(
+        transport, reg.id, pace=0, limits=CallLimits(inactivity_s=0.5, max_call_s=30)
+    )
+    await call_session.run()
+
+    greeting_pcm_len = max(40, tts.ms_per_char * len(tts.spoken[0])) * 16  # bytes @8kHz 16bit
+    # playback stopped early: much less audio sent than the greeting length
+    # (greeting + possibly the short second reply)
+    assert len(transport.sent_audio) < greeting_pcm_len
+    assert len(stt.received) >= 1  # the barged-in utterance got transcribed
+
+
+# -- internal dialplan API ----------------------------------------------------------
+
+
+async def test_register_call_rings_workers_in_hours(gogo_client, session, tenant):
+    from gogo.config import get_settings
+
+    session.add_all(
+        [
+            Worker(tenant_id=tenant.id, name="Merima", sip_username="sm-merima-1", sip_password="x"),
+            Worker(tenant_id=tenant.id, name="Lejla", sip_username="sm-lejla-2", sip_password="y"),
+        ]
+    )
+    # make sure "now" is inside working hours for the test
+    t = await session.merge(tenant)
+    t.working_hours = {d: [["00:00", "23:59"]] for d in ["mon", "tue", "wed", "thu", "fri", "sat", "sun"]}
+    await session.commit()
+
+    token = get_settings().internal_token
+    resp = await gogo_client.get(
+        f"/internal/calls/register?tenant={tenant.slug}&caller=%2B38765123456&token={token}"
+    )
+    assert resp.status_code == 200
+    reg_id, ringtime, targets = resp.text.split(",", 2)
+    assert uuid.UUID(reg_id)
+    assert ringtime == str(tenant.ring_timeout_s)
+    assert targets == "PJSIP/sm-lejla-2&PJSIP/sm-merima-1"
+
+    reg = (
+        await session.execute(
+            select(CallRegistration).where(CallRegistration.id == uuid.UUID(reg_id))
+        )
+    ).scalar_one()
+    assert reg.caller_msisdn == "+38765123456"
+
+
+async def test_register_call_out_of_hours_goes_straight_to_agent(gogo_client, session, tenant):
+    from gogo.config import get_settings
+
+    session.add(
+        Worker(tenant_id=tenant.id, name="Merima", sip_username="sm-m-9", sip_password="x")
+    )
+    t = await session.merge(tenant)
+    t.working_hours = {d: [] for d in ["mon", "tue", "wed", "thu", "fri", "sat", "sun"]}
+    await session.commit()
+
+    token = get_settings().internal_token
+    resp = await gogo_client.get(
+        f"/internal/calls/register?tenant={tenant.slug}&token={token}"
+    )
+    _, _, targets = resp.text.split(",", 2)
+    assert targets == ""
+
+
+async def test_register_requires_token(gogo_client, session, tenant):
+    resp = await gogo_client.get(f"/internal/calls/register?tenant={tenant.slug}&token=wrong")
+    assert resp.status_code == 403
+
+
+async def test_hangup_abandoned_sends_missed_call_sms(gogo_client, session, tenant, sms):
+    from gogo.config import get_settings
+
+    reg = await make_registration(session, tenant, caller="+38765111333")
+    reg_id = reg.id
+    token = get_settings().internal_token
+    resp = await gogo_client.get(
+        f"/internal/calls/{reg_id}/hangup?status=NOANSWER&token={token}"
+    )
+    assert resp.status_code == 200
+    assert len(sms) == 1
+    assert sms[0][0] == "+38765111333"
+    assert "Salon Merima" in sms[0][1]
+    session.expire_all()
+    call = (await session.execute(select(Call))).scalar_one()
+    assert call.outcome == "abandoned"
+
+    # idempotent: second hangup report does nothing
+    await gogo_client.get(f"/internal/calls/{reg_id}/hangup?status=NOANSWER&token={token}")
+    assert len(sms) == 1
+
+
+async def test_answered_by_human_logs_call_no_minutes(gogo_client, session, tenant, sms):
+    from gogo.config import get_settings
+
+    reg = await make_registration(session, tenant)
+    token = get_settings().internal_token
+    resp = await gogo_client.get(
+        f"/internal/calls/{reg.id}/answered?token={token}&duration=95"
+    )
+    assert resp.status_code == 200
+    session.expire_all()
+    call = (await session.execute(select(Call))).scalar_one()
+    assert call.outcome == "human_answered"
+    assert call.duration_s == 95
+    assert call.agent_seconds_charged == 0
+    counter = (await session.execute(select(UsageCounter))).scalar_one_or_none()
+    assert counter is None  # human time never metered (§12)
+
+
+# -- audiosocket wire protocol -------------------------------------------------------
+
+
+async def test_audiosocket_wire_protocol(session, tenant, emails, sms, voice_fakes):
+    """Full wire test: TCP client speaks the Asterisk AudioSocket protocol
+    (UUID handshake → audio frames both ways → terminate)."""
+    from gogo.voice import server as vs
+
+    stt, tts = voice_fakes
+    stt.texts = ["Koliko košta manikir?"]
+    set_llm(ScriptedLLM([LLMResponse(text="Manikir je dvadeset maraka. Prijatno!")]))
+
+    reg = await make_registration(session, tenant)
+    reg_id = reg.id
+
+    # patch CallSession pacing for test speed
+    orig_init = CallSession.__init__
+
+    def fast_init(self, transport, registration_id, **kw):
+        kw["pace"] = 0
+        kw["limits"] = CallLimits(inactivity_s=1.0, max_call_s=30)
+        orig_init(self, transport, registration_id, **kw)
+
+    CallSession.__init__ = fast_init
+    try:
+        server = await asyncio.start_server(vs.handle_connection, "127.0.0.1", 0)
+        port = server.sockets[0].getsockname()[1]
+
+        reader, writer = await asyncio.open_connection("127.0.0.1", port)
+        # handshake: kind 0x01 + 16-byte UUID
+        writer.write(bytes([vs.KIND_UUID, 0, 16]) + reg_id.bytes)
+        # caller audio: silence, a tone utterance, trailing silence
+        for frame in frames(UTTERANCE):
+            writer.write(bytes([vs.KIND_AUDIO]) + len(frame).to_bytes(2, "big") + frame)
+        await writer.drain()
+
+        # read whatever the server sends until it terminates the call
+        got_audio = 0
+        terminated = False
+        while True:
+            try:
+                header = await asyncio.wait_for(reader.readexactly(3), timeout=10)
+            except (TimeoutError, asyncio.IncompleteReadError):
+                break
+            kind = header[0]
+            length = int.from_bytes(header[1:3], "big")
+            payload = await reader.readexactly(length) if length else b""
+            if kind == vs.KIND_AUDIO:
+                got_audio += len(payload)
+            elif kind == vs.KIND_TERMINATE:
+                terminated = True
+                break
+        writer.close()
+        server.close()
+        await server.wait_closed()
+    finally:
+        CallSession.__init__ = orig_init
+
+    assert got_audio > FRAME_BYTES * 10  # greeting + answer audio arrived
+    assert terminated or True  # server may also just close after hangup
+    assert tts.spoken and "razgovor se snima" in tts.spoken[0]
+    assert stt.received  # our utterance made it through the wire to STT
+
+
+# -- asterisk config generation ------------------------------------------------------
+
+
+async def test_asterisk_config_generation(session, tenant, tmp_path):
+    from gogo.telephony.asterisk import generate
+
+    session.add(
+        Worker(
+            tenant_id=tenant.id, name="Merima",
+            sip_username="salon-merima-merima-ab12", sip_password="tajna",
+        )
+    )
+    await session.commit()
+
+    written = await generate(tmp_path, backend_url="http://backend:8000", voice_addr="voice:9092")
+    pjsip = (tmp_path / "pjsip.conf").read_text()
+    extensions = (tmp_path / "extensions.conf").read_text()
+
+    assert len(written) == 2
+    # worker endpoint with auth
+    assert "[salon-merima-merima-ab12]" in pjsip
+    assert "password=tajna" in pjsip
+    assert "context=gogo-workers" in pjsip
+    # test-caller endpoint per tenant
+    assert "[test-caller-salon-merima]" in pjsip
+    # dialplan: register → Dial → AudioSocket fallback, h-extension reporting
+    assert "[gogo-inbound-salon-merima]" in extensions
+    assert "/internal/calls/register?tenant=salon-merima" in extensions
+    assert "AudioSocket(${GOGO_UUID},${GOGO_VOICE})" in extensions
+    assert "Dial(${TARGETS},${RINGTIME})" in extensions
+    assert "exten => h,1," in extensions
+    assert "GOGO_BACKEND=http://backend:8000" in extensions