Pocket-TTS-100M-ONNX-INT8

Pocket TTS 100M ONNX INT8

CPU-oriented, streaming ONNX export of Kyutai Pocket TTS, packaged for on-device voice assistants. The 120.3 MiB bundle splits the stateful model into five graphs so a runtime can generate and play one 80 ms audio frame before computing the next one.

This repository deliberately uses Kyutai's public pocket-tts-without-voice-cloning checkpoint and bakes in its alba preset. It is English-only and does not support arbitrary voice cloning.

Files

All audio is mono Float32 PCM at 24 kHz. Each decoder invocation returns 1,920 samples (80 ms). manifest.json records exact source/exporter revisions, opsets, byte sizes and SHA-256 checksums.

Download

huggingface-cli download soniqo/Pocket-TTS-100M-ONNX-INT8 \
  --revision v1.0.0 --local-dir pocket-tts-onnx

speech-core usage

soniqo/speech-core provides a C++17 ONNX Runtime backend that performs true frame interleaving:

#include <speech_core/models/onnx_pocket_tts.h>

speech_core::PocketTtsConfig config;
config.intra_threads = 2;
config.flow_steps = 4;

speech_core::OnnxPocketTts tts("/models/pocket-tts-onnx", config);
tts.synthesize("Hello world.", "en",
    [](const float* samples, size_t count, bool final) {
        // Every non-final callback is one playable 80 ms frame.
    });

The five graphs are also compatible with sherpa-onnx. Its current offline Pocket wrapper buffers the latent sequence before decoder callbacks; use speech-core when first-frame latency is the primary requirement.

Galaxy S23 Ultra results

Measured on a Samsung SM-S918B, Snapdragon 8 Gen 2, Android 16, ONNX Runtime 1.27 CPU, two threads, four flow steps:

The native backend emitted exactly one 1,920-sample frame per callback. On Windows, native speech-core and sherpa-onnx using the same ONNX Runtime build produced the same sample count and Float32 waveform hash for the fixed-seed Hello world. test.

Round-trip intelligibility

• Deployed Parakeet-EOU recognizer: three Pocket seeds, 138 syntheses, 6.79% micro WER, 2.91% CER, 103/138 exact, zero empty output and zero EOS failure.
• Assistant-response subset: 1.96% WER and 0.06% CER across 60 syntheses.
• Independent Faster-Whisper Large-v3 pass over all 46 untouched seed-123 device WAVs: 1.43% WER, 0.40% CER and 42/46 exact after standard English normalization. Numbers and technical responses were exact.

Round-trip ASR measures intelligibility, not naturalness or listener preference. Human listening remains the appropriate gate for voice quality and prosody.

Reproducibility

The export recipe is maintained in soniqo/speech-models. It pins:

• source checkpoint kyutai/pocket-tts-without-voice-cloning at d4fdd22ae8c8e1cb3634e150ebeff1dab2d16df3;
• ONNX exporter https://github.com/csukuangfj/pocket-tts-onnx-export.git at f075c00bf4bbfbb081a11fd99abbf39df3849e0c;
• all Python/export dependencies in pyproject.toml.

A clean repack reproduced every non-manifest artifact byte-for-byte and all five graphs passed full ONNX validation.

Limitations

• English only and fixed alba voice; no arbitrary voice cloning.
• CPU optimized. NNAPI/QNN execution is not validated for these recurrent graphs.
• Dates, isolated digit sequences, uncommon proper names and tongue-twisters should be normalized or covered by product-specific regression tests.
• This model card does not claim a human MOS or preference result.

License and attribution

The model checkpoint is CC BY 4.0. The baked-in voice was performed by Alba MacKenna and is also released under CC BY 4.0; retain that attribution in redistributed bundles and product notices. Voice source: https://huggingface.co/kyutai/tts-voices#alba-mackenna.

Kyutai's implementation and the ONNX exporter code are MIT licensed. See LICENSE, the upstream repositories, and manifest.json for the exact source and artifact provenance.