Voice Triage POC

Local-first voice triage scaffold for a dev PC. You can run either:

• a local web app (voice_triage web) with start/stop mic controls and chat-style transcript/response display
• a terminal demo (voice_triage demo) for quick CLI checks

What is implemented

• Python 3.11+ project managed with uv
• Local web server + UI: uv run voice_triage web
• REST API server (UI-independent): uv run voice_triage api
• Browser microphone capture with Start/Stop controls (no push-to-talk prompt)
• Per-turn ASR using local whisper.cpp
• Multi-turn conversation engine with routing and move-home state progression
• Heuristic extractor (intent + UK postcode + basic fields)
• Minimal RAG with sqlite-backed chunk index from ./kb
• Optional BYO inference backend via REST (VOICE_TRIAGE_INFERENCE_BACKEND=byo)
• Workflow handlers (move_home, electoral_register, council_tax stubs)
• SQLite persistence for turns (./data/voice_triage.db)
• Tests with pytest
• Lint/type tooling (ruff, mypy) + pre-commit + GitHub Actions CI
• Developer references for REST API, BYO inference, telephony, and MCP under docs/

Quickstart

1. Install Python 3.11+ and uv.
2. Create and activate a project-local virtual environment.

PowerShell:

uv venv .venv
.\.venv\Scripts\Activate.ps1

Bash:

uv venv .venv
source .venv/bin/activate

1. Sync dependencies into .venv:

uv sync --dev

1. Bake runtime paths into .venv/.env (recommended):

.\scripts\configure_venv_env.ps1 `
  -WhisperBin ".venv/tools/whispercpp/Release/whisper-cli.exe" `
  -WhisperModel ".venv/tools/whispercpp/models/ggml-base.en.bin" `
  -WhisperUseGpu $true `
  -WhisperGpuLayers 60 `
  -WhisperThreads 6 `
  -WhisperTimeoutSeconds 45 `
  -InferenceBackend "local" `
  -ByoInferenceUrl "" `
  -ByoApiStyle "generic" `
  -ByoModel "" `
  -ByoApiKey "" `
  -ByoSystemPrompt "" `
  -PiperBin ".venv/Scripts/piper.exe" `
  -PiperModel ".venv/tools/piper/models/en_GB-alba-medium.onnx" `
  -PiperDefaultVoiceId "en_GB-alba-medium" `
  -PiperTimeoutSeconds 30 `
  -WebHost "0.0.0.0" `
  -WebPort 8443 `
  -SslCertFile ".venv/certs/dev-cert.pem" `
  -SslKeyFile ".venv/certs/dev-key.pem"

This writes .venv/.env (project-local), and the app auto-loads it.

1. Generate TLS certs (required for LAN browser mic access):

.\scripts\generate_dev_tls_cert.ps1

By default this includes localhost, 127.0.0.1, and detected LAN IPv4 addresses.

1. (Optional) Keep tools inside .venv for fully self-contained local runtime:

• whisper.cpp binary: .venv/tools/whispercpp/whisper-cli.exe (or .venv/tools/whispercpp/main.exe)
• whisper model: .venv/tools/whispercpp/models/ggml-base.en.bin
• piper binary: .venv/tools/piper/piper.exe
• piper model: .venv/tools/piper/models/en_GB-northern_english_male-medium.onnx
  • any additional *.onnx files in .venv/tools/piper/models appear in the web voice dropdown
  • default web voice can be set with PIPER_DEFAULT_VOICE_ID (defaults to en_GB-alba-medium)

1. (Optional) Add local knowledge base files under ./kb/*.md or ./kb/*.txt.

Reindex knowledge base on demand (no restart required):

uv run voice_triage reindex

Run local website

Start server:

uv run voice_triage web --host 127.0.0.1 --port 8000

Then open:

http://127.0.0.1:8000

The web UI now uses the versioned REST surface under /api/v1/*.

For LAN access:

uv run voice_triage web --host 0.0.0.0 --port 8443 --ssl-certfile .venv/certs/dev-cert.pem --ssl-keyfile .venv/certs/dev-key.pem

Then browse from another device on your LAN:

https://<your-pc-lan-ip>:8443

Run REST API only

uv run voice_triage api --host 127.0.0.1 --port 8000 --no-ssl

Useful endpoints:

• POST /api/v1/session
• POST /api/v1/session/{session_id}/turn (audio upload)
• POST /api/v1/session/{session_id}/turn/text (pre-transcribed text)
• GET /api/v1/voices
• GET /api/v1/config (client runtime config, including VAD tuning)
• POST /api/v1/reindex (rebuild RAG index from current ./kb)
• POST /api/v1/session/{session_id}/voice
• GET /api/v1/tts/{audio_id}

Important for browser microphone access:

• Many browsers block getUserMedia on plain http://<lan-ip>.
• localhost is usually allowed without HTTPS, but LAN IPs usually require HTTPS.
• For LAN clients, use HTTPS with a cert that includes your LAN IP in SANs.

Web flow:

1. Click Start Listening
2. Speak naturally
3. The browser VAD auto-detects end-of-turn silence and sends the turn
4. See:
   • user transcript (what ASR heard)
   • assistant response
   • assistant audio playback (Piper)
   • selectable Piper voice (dropdown)
5. The app keeps listening for the next turn; click Stop Listening to end hands-free mode

CLI demo (legacy)

uv run voice_triage demo

Optional MCP server

Run MCP stdio server:

uv run voice_triage mcp

MCP also exposes reindex_kb for on-demand knowledge base refresh.

If MCP SDK is missing, install optional dependency:

uv add --optional mcp "mcp>=1.0.0"

Common commands

make sync
make format
make lint
make typecheck
make docstrings
make test
make telephony-contract
make telephony-smoke-local
# with custom base URL:
# make telephony-smoke-remote BASE_URL=https://your-public-host
make demo
make web
make api
make web-lan
make stop-web
make cert-dev
make web-ssl

Developer references

• REST API reference: docs/API_REFERENCE.md
• BYO inference reference: docs/BYO_INFERENCE_REFERENCE.md
• Telephony provider how-tos: docs/TELEPHONY_REFERENCE.md
• Telephony guide index: docs/telephony/README.md
• Telephony integration status: docs/TELEPHONY_INTEGRATIONS.md
• MCP reference: docs/MCP_REFERENCE.md

Audio prerequisites (brief)

• Ubuntu/Debian: sudo apt-get install -y portaudio19-dev
• macOS: brew install portaudio
• Windows: no extra package is usually needed when using standard Python wheels.

CUDA acceleration for whisper.cpp

For faster STT turn latency on NVIDIA GPUs, use a CUDA-enabled whisper.cpp build and set:

WHISPERCPP_USE_GPU=1
WHISPERCPP_GPU_LAYERS=60
WHISPERCPP_THREADS=6
WHISPERCPP_TIMEOUT_SECONDS=45

Notes:

• Increase WHISPERCPP_GPU_LAYERS toward 99 if VRAM allows.
• If your build fails with GPU flags, set WHISPERCPP_USE_GPU=0.
• You can pass additional whisper CLI flags via WHISPERCPP_EXTRA_ARGS in .venv/.env.

BYO inference backend

Set:

VOICE_TRIAGE_INFERENCE_BACKEND=byo
VOICE_TRIAGE_BYO_INFERENCE_URL=http://localhost:9000/infer
VOICE_TRIAGE_BYO_INFERENCE_TIMEOUT_SECONDS=12
VOICE_TRIAGE_BYO_API_STYLE=generic

Expected BYO endpoint request payload:

{"query":"How do I order a garden waste bin?"}

Expected response payload:

{"answer":"...", "metadata":{"provider":"my-model"}}

Ollama (OpenAI-compatible) example:

VOICE_TRIAGE_INFERENCE_BACKEND=byo
VOICE_TRIAGE_BYO_API_STYLE=openai
VOICE_TRIAGE_BYO_INFERENCE_URL=http://127.0.0.1:11434/v1/chat/completions
VOICE_TRIAGE_BYO_MODEL=llama3.1:8b
VOICE_TRIAGE_BYO_INFERENCE_TIMEOUT_SECONDS=20
VOICE_TRIAGE_BYO_API_KEY=

When BYO is unavailable, local RAG fallback is used automatically.

Notes

• If ./kb has no indexed chunks, RAG answers with a clear fallback message asking the user to rephrase or ask about supported council service topics.
• TTS normalizes UK currency values for more natural speech (for example £33.50 becomes 33 pounds and 50 pence).
• Browser mic recording requires microphone permission for localhost.
• If Piper is not configured, text responses still work and UI shows a TTS error message.
• If LAN clients cannot connect, allow inbound TCP 8443 in Windows Firewall.
• If Ctrl+C fails in one terminal, run uv run voice_triage stop-web from another terminal.