fastrtc-pocket-tts

A FastRTC-compatible Text-to-Speech (TTS) adapter for Kyutai's ultra-low latency Pocket TTS.

Pocket TTS is a highly optimized, CPU-bound model that requires no GPU and generates speech up to 6x faster than real-time (~200ms time-to-first-audio).

Features

• ⚡ Ultra-Low Latency: ~200ms to the first audio chunk.
• 💻 CPU-based: Runs effortlessly on Mac, Windows, and Linux without CUDA/Metal dependencies.
• 🧵 Non-Blocking: Implements threaded async generation to keep your WebRTC/FastAPI event loops responsive.
• 🗣️ Instant Voice Cloning: Supply a .wav file, and it clones the voice instantly.

Installation

Install via pip:

pip install fastrtc-pocket-tts

This package requires Python 3.10+ and automatically installs fastrtc and pocket-tts as dependencies.

Quickstart

Run a local conversational AI that responds to you in real-time using your computer's microphone.

from fastrtc import Stream, ReplyOnPause
from fastrtc_pocket_tts import PocketTTS, PocketVoice

# Initialize the model using a built-in voice
tts_engine = PocketTTS(default_voice=PocketVoice.ALBA)

def voice_handler(audio):
    # In a real app, you would pass `audio` to an STT and LLM here.
    # For this example, we'll just respond with static text.
    response_text = "I am speaking to you directly from your CPU."
    
    # Yield the audio stream chunks back to FastRTC
    for chunk in tts_engine.stream_tts_sync(response_text):
        yield chunk

# Mount it on a FastRTC stream
stream = Stream(ReplyOnPause(voice_handler), modality="audio", mode="send-receive")

# Launch the interactive WebRTC UI
stream.ui.launch()

Voice Cloning

Pocket TTS allows zero-shot voice cloning. Simply provide a valid path to a local audio file (.wav, .mp3) when initializing the model or passing options.

from fastrtc_pocket_tts import PocketTTS

# The model will automatically extract the voice embedding from the audio file
tts_engine = PocketTTS(default_voice="./path/to/my_voice_sample.wav")


Available Built-In Voices
If you don't want to use voice cloning, Kyutai provides 8 pre-trained English default voices:
alba, marius, javert, jean, fantine, cosette, eponine, azelma.
You can access them safely using the PocketVoice Enum:
from fastrtc_pocket_tts import PocketVoice

voice = PocketVoice.JAVERT

License

MIT License.