Pocket TTS 100M
Pocket TTS 100M is the low-latency, on-device English text-to-speech backend in speech-core. The public ONNX bundle uses the fixed Alba voice and streams audio as soon as each acoustic frame is ready.
Inference is genuinely interleaved: one autoregressive latent is decoded and emitted before generation of the next latent. Each non-final callback carries 1,920 Float32 samples — 80 ms at 24 kHz — avoiding the fixed full-graph floor of buffered TTS exports.
Published bundle
| Item | Details |
|---|---|
| Model | soniqo/Pocket-TTS-100M-ONNX-INT8 |
| Release | v1.0.0 · ad5d3fae3fb2f10024583708597afca8337c094c |
| Bundle size | 120.3 MiB |
| Runtime | ONNX Runtime · C++17 · dynamic INT8 + FP32 |
| Output | 24 kHz mono Float32 PCM · 1,920 samples / 80 ms per frame |
| Language / voice | English · fixed Alba voice |
Download and verify
The speech-core downloader pins v1.0.0 and verifies every published file by SHA-256.
bash scripts/download_pocket_tts_onnx.sh /models/pocket-tts-onnx
C++ API
#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 length, bool is_final) {
// Each non-final callback is 1,920 Float32 samples:
// 80 ms at 24 kHz.
});
Streaming contract
The callback receives one 80 ms frame at a time; the final callback marks end of stream. last_metrics() reports conditioning, first-audio and total latency, frame and sample counts, EOS state, cancellation, and the effective seed. cancel() stops an active synthesis.
Galaxy S23 Ultra benchmark
A clean rerun of the public v1.0.0 bundle used ONNX Runtime 1.27, two CPU threads, and a Galaxy S23 Ultra. These are direct engine measurements, not UI or complete voice-pipeline timings.
| Metric | Result |
|---|---|
| TTFA p50 / p95 | 127.7 / 133.7 ms |
| Total latency p50 / p95 | 467.5 / 485.3 ms |
| Median RTF | 0.4495 |
| Peak RSS | 373.7 MiB |
Round-trip intelligibility
The device gate synthesized 46 command and assistant phrases, resampled the audio through the production 24-to-16 kHz path, and transcribed it with the deployed Parakeet-EOU recognizer.
| Metric | Result |
|---|---|
| Corpus | 46 phrases |
| Exact transcripts | 31 / 46 |
| WER / CER | 8.00% / 3.47% |
| Empty audio / transcript / EOS failures | 0 / 0 / 0 |
Round-trip ASR measures intelligibility, not naturalness or listener preference; it complements listening tests.
Integration status
The C++17 ONNX backend is merged in speech-core and builds for Windows, Linux, and Android. The current speech-android SDK/JNI path still selects Kokoro by default; Kotlin model selection, download wiring, and SDK exposure are follow-up work. This release does not include an Apple/CoreML backend.
Limits and license
- English only; the public checkpoint bakes in the fixed Alba voice and does not support arbitrary voice cloning.
- The model and voice are CC BY 4.0. Keep the Alba MacKenna attribution included in the bundle manifest.
- The recurrent graphs are CPU-oriented. Hardware acceleration is opt-in because Android execution-provider support varies by device and ONNX Runtime build.
- The speech-core source is Apache 2.0; the model bundle retains its own CC BY 4.0 terms.