pipecat-maya 0.1.0

Maya Research TTS integration for Pipecat

pipecat-maya

Maya Research TTS integration for Pipecat.

Maya 2 is a text-to-speech model built for Indian languages. It offers two voices — Ananya (female) and Arjun (male) — and both support all ten available languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu.

This package provides MayaHttpTTSService, a drop-in Pipecat TTS service that streams audio from Maya's HTTP API.

Maintained by the Maya Research team. Tested with Pipecat v1.6.0, verified against the live Maya 2 API.

Installation

uv add pipecat-maya
# or
pip install pipecat-maya

pipecat-ai is pulled in automatically. The service itself needs no extra dependencies beyond aiohttp.

Prerequisites

You need a Maya API key. Contact [e-mail], [e-mail], or [e-mail] to get one.

export MAYA_API_KEY=...

Usage

import os

from pipecat.transcriptions.language import Language
from pipecat_maya import MayaHttpTTSService

tts = MayaHttpTTSService(
    api_key=os.environ["MAYA_API_KEY"],
    settings=MayaHttpTTSService.Settings(
        voice="Ananya",  # or "Arjun"
        language=Language.HI,
        region="IN",  # or "US"
    ),
)

Then drop it into a pipeline exactly like any built-in Pipecat TTS service:

pipeline = Pipeline(
    [
        transport.input(),
        stt,
        user_aggregator,
        llm,
        tts,  # <-- Maya
        transport.output(),
        assistant_aggregator,
    ]
)

Configuration

Constructor parameters:

Parameter	Required	Default	Description
api_key	Yes	—	Maya API key.
aiohttp_session	No	None	Reuse an existing aiohttp.ClientSession. If omitted, the service creates and closes its own.
base_url	No	https://tts.mayaresearch.ai	Override for self-hosted or staging endpoints.
sample_rate	No	pipeline default	Output rate in Hz. Maya synthesizes at 24 kHz; anything else is resampled.
settings	No	see below	Runtime-updatable settings.

Runtime settings (MayaHttpTTSService.Settings), all changeable mid-pipeline via TTSUpdateSettingsFrame:

Setting	Default	Values
voice	"Ananya"	"Ananya", "Arjun"
language	None	Any Language enum; the ten Indian languages listed above are verified.
region	"IN"	"IN", "US"

model is unused — Maya 2 exposes a single model.

Maya validates language strictly: an unsupported code is rejected with 400 invalid 'language'. So if you configure a language Maya doesn't support (English, for example), the service logs a warning and omits the field, letting Maya fall back to its own default — you still get audio instead of a failed turn.

Supported codes: hi, bn, gu, kn, ml, mr, or, pa, ta, te.

Examples

Verify your key without setting up a full pipeline — synthesizes one Hindi sentence to maya-out.wav:

MAYA_API_KEY=... uv run --with pipecat-maya example/synthesize_to_wav.py

Full voice agent (Deepgram STT + OpenAI LLM + Maya TTS over WebRTC). Needs DEEPGRAM_API_KEY and OPENAI_API_KEY in addition to MAYA_API_KEY:

uv run --with "pipecat-maya[example]" example/maya_voice_agent.py

Then open the printed URL in your browser.

Local mic/speaker agent with latency instrumentation — talks through your laptop's mic and speakers, no browser or transport server. Prints per-service TTFB, an end-to-end latency breakdown, transcripts, and LLM activity. Needs SONIOX_API_KEY and OPENAI_API_KEY alongside MAYA_API_KEY:

uv run --with "pipecat-maya[local-agent]" example/local_voice_agent_metrics.py

Set MAYA_VOICE (Ananya/Arjun) and MAYA_LANGUAGE (hi, ta, ...) to switch voice and language.

How it works

Maya's POST /v1/tts streams raw PCM — 16-bit signed little-endian, mono, 24 kHz, no file header. The service reads the response with iter_chunked and pushes TTSAudioRawFrames as bytes arrive rather than buffering the full utterance, so time-to-first-byte stays low. Pipecat's _stream_audio_frames_from_iterator handles resampling to the pipeline rate and keeps chunks sample-aligned.

The HTTP session is reused across requests to avoid repeated TLS handshakes.

Rate limits

Maya's defaults are 120 requests/minute and 16 concurrent requests. Exceeding either returns a 429, which surfaces as an ErrorFrame with the status and body.

Development

uv sync --extra dev
uv run pytest
uv run ruff format . && uv run ruff check .

License

BSD 2-Clause. See LICENSE.