Pocket TTS Complete Documentation

Generate speech from text using Kyutai Pocket TTS - lightweight, CPU-friendly, streaming TTS with voice cloning. English only. ~6x real-time on M4 MacBook Air.

Pocket TTS

Lightweight CPU-friendly text-to-speech with voice cloning. No GPU required.

When to Use

• Generating speech from text on CPU without GPU
• Voice cloning from audio samples
• Streaming audio generation (low latency)
• Local TTS without API dependencies
• Real-time speech synthesis (~6x faster than real-time)

---

Key Features

• 100M parameters - Small, efficient model
• CPU-optimized - No GPU needed, uses only 2 cores
• ~6x real-time - Fast generation on modern CPUs
• ~200ms latency - To first audio chunk (streaming)
• Voice cloning - From 3-10s audio samples
• 24kHz mono WAV - High-quality output
• English only - More languages planned

---

Installation

pip install pocket-tts
# or
uv add pocket-tts

---

CLI Commands

Generate Speech

# Basic generation (default voice)
pocket-tts generate --text "Hello world"

# Custom voice (local file, URL, or safetensors)
pocket-tts generate --voice ./my_voice.wav
pocket-tts generate --voice "hf://kyutai/tts-voices/alba-mackenna/casual.wav"
pocket-tts generate --voice ./voice.safetensors

# Quality tuning
pocket-tts generate --temperature 0.7 --lsd-decode-steps 3

See docs/generate.md for full CLI reference.

Start Web Server

# Start FastAPI server with web UI
pocket-tts serve

# Custom host/port
pocket-tts serve --host localhost --port 8080

See docs/serve.md for server options.

Export Voice Embeddings

Convert audio files to .safetensors for faster loading:

# Single file
pocket-tts export-voice voice.mp3 voice.safetensors

# Batch conversion
pocket-tts export-voice voices/ embeddings/ --truncate

See docs/export_voice.md for export options.

---

Python API

Basic Usage

from pocket_tts import TTSModel
import scipy.io.wavfile

# Load model
model = TTSModel.load_model()

# Get voice state
voice = model.get_state_for_audio_prompt(
    "hf://kyutai/tts-voices/alba-mackenna/casual.wav"
)

# Generate audio
audio = model.generate_audio(voice, "Hello world!")

# Save
scipy.io.wavfile.write("output.wav", model.sample_rate, audio.numpy())

Load Model

model = TTSModel.load_model(
    config="b6369a24",       # Model variant
    temp=0.7,                # Temperature (0.5-1.0)
    lsd_decode_steps=1,      # Generation steps (1-5)
    eos_threshold=-4.0       # End-of-sequence threshold
)

Voice State

# From audio file/URL
voice = model.get_state_for_audio_prompt("./voice.wav")
voice = model.get_state_for_audio_prompt("hf://kyutai/tts-voices/alba-mackenna/casual.wav")

# From safetensors (fast loading)
voice = model.get_state_for_audio_prompt("./voice.safetensors")

Streaming Generation

# Stream audio chunks
for chunk in model.generate_audio_stream(voice, "Long text..."):
    # Process/save/play each chunk as generated
    print(f"Chunk: {chunk.shape[0]} samples")

Multi-Voice Management

# Preload multiple voices
voices = {
    "casual": model.get_state_for_audio_prompt("hf://kyutai/tts-voices/alba-mackenna/casual.wav"),
    "announcer": model.get_state_for_audio_prompt("./announcer.safetensors"),
}

# Use different voices
audio1 = model.generate_audio(voices["casual"], "Hey there!")
audio2 = model.generate_audio(voices["announcer"], "Breaking news!")

See docs/python-api.md for complete API reference.

---

Available Voices

Pre-made voices from hf://kyutai/tts-voices/:

• alba-mackenna/casual.wav (default, female)
• jessica-jian/casual.wav (female)
• voice-donations/Selfie.wav (male, marius)
• voice-donations/Butter.wav (male, javert)
• ears/p010/freeform_speech_01.wav (male, jean)
• vctk/p244_023.wav (female, fantine)
• vctk/p262_023.wav (female, eponine)
• vctk/p303_023.wav (female, azelma)

Or clone any voice from your own audio samples.

---

Voice Cloning Tips

• Clean audio - Remove background noise (use Adobe Podcast Enhance)
• Length - 3-10 seconds of speech is ideal
• Quality - Input quality affects output quality
• Format - WAV, MP3, or any common audio format supported

---

Performance Tips

• CPU-only - GPU provides no speedup (model too small, batch size 1)
• 2 cores - Uses only 2 CPU cores efficiently
• Streaming - Low latency (<200ms to first chunk)
• Safetensors - Pre-process voices to .safetensors for instant loading

---

Output Format

All commands output WAV files:

• Sample rate: 24 kHz
• Channels: Mono
• Bit depth: 16-bit PCM

---

Links

• GitHub
• Tech Report
• Paper (arXiv)
• HuggingFace Model
• Voice Repository
• Live Demo