Pocket TTS Voice Cloning for LiteRT

This repository contains 22 Pocket TTS model packs covering 26 languages, exported to LiteRT for on-device speech synthesis. Each pack can synthesize speech in a voice conditioned on a short reference WAV. The output is mono audio at 24 kHz.

The graphs store weights in FP16 and use FP32 inputs and outputs. The export notebook checks each graph on LiteRT CPU. Performance depends on the device and checkpoint. The Korean and Welsh packs use larger 24-layer models.

Languages

Checkpoints Languages
Official Kyutai English, Dutch, French, German, Italian, Portuguese, Spanish
Community language packs Czech, Estonian, Greek, Gujarati, Indonesian, Japanese, Korean, Mandarin Chinese, Persian, Polish, Turkish, Welsh, Wolof
Shared Indic multilingual pack English, Hindi, Kannada, Malayalam, Marathi, Tamil, Telugu

The January English checkpoint and the newer English checkpoint are available as separate packs. The six Indic language options share one download.

Choose a pack

Open pocket-tts-index.json to find the pack for your target language. Download the files listed in that pack's pocket-tts.json. The manifest also records tensor shapes, text preprocessing rules, source checkpoint information, and SHA-256 checksums.

Each pack includes four .tflite graphs for reference-audio encoding, autoregressive generation, and the two Mimi decoding stages. It also includes the tokenizer and the weights used by the host code. The host tokenizes the text, maintains the generation cache, and assembles the audio. The accompanying Colab notebook shows how to run the complete pipeline with a reference WAV.

Credits and licenses

Pocket TTS and the Mimi codec were developed by Kyutai. The conversion builds on john-rocky’s LiteRT exporter. See the README.md inside each pack for its original checkpoint, authors, and license. Use reference recordings with the speaker’s permission.

Downloads last month
345
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support