Pipecat

Pipecat builds voice agents as a pipeline of services. Bodhi provides the speech-to-text stage through the Python SDK, so transcription behaves like any other Pipecat STT service.

Install

pip install "bodhi-api-sdk[stt]"

Needs pipecat-ai 1.4 or newer. Verified against 1.4.0, 1.8.1 and 1.11.0.

Add it to a pipeline

import os

from pipecat.pipeline.pipeline import Pipeline

from bodhi.integrations.pipecat_stt import BodhiHotword, BodhiSTTService

stt = BodhiSTTService(
    api_key=os.environ["BODHI_API_KEY"],
    model="hi-banking-v2-8khz",
    settings=BodhiSTTService.Settings(
        parse_number=True,
        endpoint_silence_duration=0.6,
        hotwords=[BodhiHotword("बजाज फिनसर्व", 2.0)],
    ),
)

pipeline = Pipeline([
    transport.input(),
    stt,
    context_aggregator.user(),
    llm,
    tts,
    transport.output(),
])

api_key and model are the only required arguments. Everything else has a default, including url, which points at wss://stt.navana.ai.

Arguments

Argument	Required	Description
`api_key`	Yes	Your platform API key, sent as the `x-api-key` header.
`model`	Yes	Any transcription model. See Models and languages.
`url`	No	Override the endpoint, usually to point at a different deployment.
`sample_rate`	No	Input rate in Hz. Defaults to the pipeline’s own.
`language`	No	Language tag for emitted frames. Defaults to the model’s prefix.
`interim_results`	No	Emit `InterimTranscriptionFrame`s from partials. Defaults to on.
`aux`	No	Attach timing and confidence metadata to each frame.
`min_confidence`	No	Drop finals below this utterance confidence. Defaults to `0.5`; pass `0` to keep everything.
`transaction_id`	No	Correlation id for the session. A fresh UUID per connection when omitted.
`settings`	No	Runtime-updatable recognition settings, below.

Settings

Pass these through BodhiSTTService.Settings. They can be updated while the agent is running.

Setting	Description
`model`	The transcription model. Changing it reconnects the session.
`language`	Language used to tag emitted frames. Presentational only, since Bodhi infers the language from the model.
`hotwords`	Phrase boosting, as `BodhiHotword(phrase, score)`. `score` is optional, and roughly `1.0` to `3.0` is useful. See Advanced features.
`parse_number`	Normalise spoken numbers, dates, and currency in the returned text.
`endpoint_silence_duration`	Trailing silence before an utterance is finalised, in seconds. Clamped to `0.44` to `1.2`, and left to the server’s own `0.44` when unset.

How it maps onto Pipecat

Partial transcripts arrive as InterimTranscriptionFrames and endpointed finals as TranscriptionFrames, which is the contract every Pipecat STT service follows. Interruption handling and turn taking work unchanged, so swapping another provider for Bodhi does not touch the rest of the pipeline.

Python SDK

Using the SDK directly, without Pipecat.

Models and languages

Every model, and how to pick one.

Advanced features

Hotwords, number parsing, endpointing.