pipecat-plugins-munsit 0.2.0

Pipecat plugin for Munsit Arabic speech-to-text and text-to-speech

pipecat-plugins-munsit

Pipecat services for Munsit: Arabic speech-to-text (live WebSocket streaming) and text-to-speech (HTTP streaming).

This package supersedes pipecat-plugins-faseeh, which is now a deprecation shim that re-exports these services.

Installation

pip install pipecat-plugins-munsit

Requires Python 3.10+ and pipecat-ai >= 1.5.0.

Speech-to-Text (live streaming)

MunsitSTTService streams audio to Munsit's live WebSocket API (wss://api.munsit.com/api/v1/listen) and pushes interim and final transcription frames. Arabic only (v1).

from pipecat_plugins_munsit import MunsitSTTService

stt = MunsitSTTService(
    api_key="your-api-key",       # or MUNSIT_API_KEY env var
    endpointing_ms=800,           # server-side silence window that ends a turn (100-5000)
    hotwords=["فصيح", "منصت"],    # custom vocabulary biasing (<=200 entries, <=40 chars)
)

Per-turn enrichment (speaker gender, sentiment)

async def on_enrichment(event: dict):
    if event["type"] == "Gender":
        print("speaker gender:", event["label"], event["score"])
    elif event["type"] == "Sentiment":
        print("turn sentiment:", event["label"], event["score"])

stt = MunsitSTTService(api_key="...", on_enrichment=on_enrichment)

Notes

• Input at 8 kHz / 16 kHz is sent as-is; other pipeline rates are resampled to 16 kHz client-side.
• Idle connections are kept open with protocol KeepAlive control frames (never silence, which would be billed as audio).
• A final with TranscriptionFrame.finalized == False is a forced split inside one long utterance (the server splits turns at ~60s of unbroken speech), not the end of the user's turn.
• smart_turn (on by default) gates end-of-turn on the server's semantic turn-completion model, reducing premature cuts; pass smart_turn=False to end turns on silence alone.

Text-to-Speech

import aiohttp
from pipecat_plugins_munsit import MunsitTTSService

async with aiohttp.ClientSession() as session:
    tts = MunsitTTSService(
        api_key="your-api-key",   # or MUNSIT_API_KEY env var
        aiohttp_session=session,
        voice_id="ar-hijazi-female-2",
        model="faseeh-v1-preview",
    )

Streams PCM16 audio at 48 kHz mono by default (sample_rate=24000 also supported). stability (0.0-1.0) and speed (0.7-1.2) are configurable at construction and runtime-updatable via TTSUpdateSettingsFrame.

Full agent example

from pipecat.pipeline.pipeline import Pipeline
from pipecat_plugins_munsit import MunsitSTTService, MunsitTTSService

pipeline = Pipeline([
    transport.input(),
    MunsitSTTService(api_key="..."),
    context_aggregator.user(),
    llm,
    MunsitTTSService(api_key="...", aiohttp_session=session),
    transport.output(),
    context_aggregator.assistant(),
])

Development

pip install -e ".[dev]"
pytest tests/ -q

Releases are published to PyPI by pushing a v* tag (see .github/workflows/publish.yml; requires the PYPI_API_TOKEN repo secret).

License

Apache-2.0