Hugging Face's logo

theoracleguy
/
pocket-tts

Pocket-TTS
Safetensors
English
Model card Files
xet
 Community

OpenVox TTS — by The Oracle Guy

A lightweight, CPU-friendly text-to-speech model integrated into OpenVox, the local AI voice synthesis app for macOS and Windows. Built for developers and creators who want fast, private TTS without cloud APIs or GPU requirements.

🔊 Try OpenVox | 🤗 Hugging Face | ▶️ YouTube

What this model powers

This model is the TTS backbone used in OpenVox, which provides:

  • Local, offline text-to-speech — no internet required after install
  • Voice cloning from a short audio sample
  • 600+ language support via OmniVoice engine
  • Audiobook generation from EPUB/PDF
  • Multi-speaker Conversations mode (up to 4 voices)
  • OpenAI-compatible local API
  • macOS (Mac App Store) and Windows apps

Main takeaways

  • Runs entirely on CPU — no GPU needed
  • Small model footprint (~100M parameters)
  • Audio streaming with ~200ms latency to first chunk
  • ~6x real-time speed on Apple Silicon (M-series)
  • Uses only 2 CPU cores
  • Python API and CLI available
  • Voice cloning via short audio prompt
  • English-first, with multilingual support in OpenVox via OmniVoice
  • Can handle arbitrarily long text inputs
  • Can run client-side in the browser (WebAssembly)

Try it without installing anything

Visit openvoxai.com to try OpenVox directly, or use the Kyutai demo at kyutai.org/pocket-tts to test the underlying model in your browser.

CLI usage

Generate audio 

uvx pocket-tts generate
# or after pip install:
pocket-tts generate

Modify voice and text:

pocket-tts generate --voice alba --text "Hello from OpenVox."

Available built-in voices: alba, marius, javert, jean, fantine, cosette, eponine, azelma

The --voice argument also accepts a local .wav file for voice cloning.

Run a local server 

pocket-tts serve

Navigate to http://localhost:8000 for the web UI. The model stays in memory between requests, so it's faster for iterating on voices and prompts.

Export a voice embedding

Processing audio for voice cloning is slow. Use export-voice to convert an audio file to a fast-loading .safetensors embedding:

pocket-tts export-voice --input my_voice.wav --output my_voice.safetensors

Python API

Install:

pip install pocket-tts

Basic usage:

from pocket_tts import TTSModel
import scipy.io.wavfile

tts_model = TTSModel.load_model()
voice_state = tts_model.get_state_for_audio_prompt("alba")
audio = tts_model.generate_audio(voice_state, "Hello from OpenVox.")
scipy.io.wavfile.write("output.wav", tts_model.sample_rate, audio.numpy())

For performance, keep load_model() and get_state_for_audio_prompt() results in memory — both are slow operations.

Fast voice loading with safetensors 

from pocket_tts import TTSModel, export_model_state, import_model_state

model = TTSModel.load_model()

# Export once
voice_state = model.get_state_for_audio_prompt("alba")
export_model_state(voice_state, "alba.safetensors")

# Load fast on subsequent runs
voice_state = import_model_state("alba.safetensors")
audio = model.generate_audio(voice_state, "Hello from OpenVox.")

OpenVox integration

This model is distributed as part of OpenVox, available on:

OpenVox wraps this model in a polished native UI with features like audiobook import, multi-speaker scripts, emotion tags via OmniVoice, and a local OpenAI-compatible API endpoint.

Pricing: free tier (5,000 characters/day) + one-time $20 Pro upgrade. No subscriptions.

In-browser implementations

Pocket TTS (the underlying model) is small enough to run in WebAssembly:

Alternative backends

Unsupported features (contributions welcome)

  • Silence/pause injection via text input
  • int8 quantization

GPU execution was tested but showed no speedup over CPU at batch size 1 with this model size.

Prohibited use

Use of this model must comply with all applicable laws and regulations. Prohibited uses include voice impersonation or cloning without explicit consent; generating misinformation, fake recordings, or fraudulent content; and producing harmful, discriminatory, or privacy-invasive audio. The Oracle Guy disclaims all liability for non-compliant use.

Credits

Model architecture based on Pocket TTS by Kyutai (CC-BY 4.0). OpenVox app and integrations by The Oracle Guy.

Downloads last month
14
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support