Pocket Tts Onnx
ONNX export of Pocket TTS for lightweight multilingual text-to-speech with zero-shot voice cloning.
Model source
- Kind
- Other
- File size
- 9.50–72.8 MB
- Base model
- kyutai/pocket-tts
- Version
- vRevision 651140c
- Updated
- Sep 21, 2026
- Downloads
- 0
- Tasks
- Audio
- Source
- Hugging Face
Source excerpt
ONNX export of Pocket TTS for lightweight multilingual text-to-speech with zero-shot voice cloning.
Sources
1 source
Sources
1 sourceVerified Sep 21
Model artifacts
35 artifacts
Model artifacts
35 artifactsSource excerpts
2 excerpts
Source excerpts
2 excerpts--- license: cc-by-4.0 language: - en - de - fr - it - pt - es library_name: pocket-tts-onnx base_model: - kyutai/pocket-tts pipeline_tag: text-to-speech tags: - tts - voice-cloning - onnx - onnxruntime --- # Pocket TTS ONNX ONNX export of [Pocket TTS](https://huggingface.co/kyutai/pocket-tts) for lightweight multilingual text-to-speech with zero-shot voice cloning. ## Model Description Pocket TTS is a compact text-to-speech model from Kyutai that supports voice cloning from short audio samples. This ONNX export provides: - **Zero-shot voice cloning** from any audio reference - **Multilingual bundles** for `english_2026-04`, `french_24l`, `german`, `german_24l`, `italian`, `italian_24l`, `portuguese`, `portuguese_24l`, `spanish`, and `spanish_24l` - **INT8 quantized models** for fast CPU inference - **Streaming support** with adaptive chunking for real-time playback - **Temperature control** for generation diversity - **Dual model architecture** for flexible flow matching ## Architecture This export uses a **dual model split** for the Flow LM: ``` flow_lm_main - Transformer/conditioner (produces conditioning vectors) flow_lm_flow - Flow network only (Euler integration for latent sampling) ``` This architecture enables: - **Temperature control**: Adjust generation diversity via noise scaling - **Variable LSD steps**: Trade off speed vs quality - **External flow loop**: Full control over the sampling process ## Performance Performance depends on the selected language bundle, precision, CPU, and `lsd_steps`. - 6-layer bundles such as `english_2026-04`, `german`, `italian`, `portuguese`, and `spanish` are much smaller and faster than the `*_24l` bundles. - INT8 bundles substantially reduce `flow_lm_main`, `flow_lm_flow`, and `mimi_decoder` size for CPU inference. - `mimi_encoder.onnx` and `text_conditioner.onnx` remain FP32. ### Optimized Inference Thread count is automatically tuned to avoid over-subscription on the small sequential matmuls in the autoregressive loop (`intra_op_num_threads=min(cpu_count, 4)`, `inter_op_num_threads=1`). This alone provides a **~2x speedup** over default ORT settings on multi-core machines. ### Decoding Strategy - **Offline (`generate`)**: Uses **threaded parallel decoding** — the mimi decoder runs in a background thread, decoding 12-frame chunks while the flow loop generates the next frames. This overlaps gene...
Source context: 0 downloads · 0 likes · Pipeline text-to-speech · Library pocket-tts-onnx · Repo deepLEARNING786/pocket-tts-onnx
