lordnyson/playback-recordings-voices

Playback Recordings voices

Voice model packs downloaded on first use by Playback Recordings, an offline Android narration app. Each pack is one uncompressed tar that runs through sherpa-onnx on the device.

Pack	Voices in the app	Size
kokoro-v1.0.tar	Heart, Michael, Bella	~360 MB
pocket-2026-09.tar	Eponine, Peter Yearsley	~160 MB

Nothing here is new model training: these are existing open models, converted or patched so that sherpa-onnx renders them the way their reference runtimes do. tools/ holds the scripts that made every changed file.

kokoro-v1.0

• model.onnx - Kokoro-82M v1.0 (Apache-2.0, hexgrad), the fp32 ONNX export from thewh1teagle/kokoro-onnx (MIT), with sherpa-onnx's model metadata copied onto it (tools/patch_kokoro_metadata.py). sherpa-onnx's own Kokoro export rebuilds the vocoder and sounds audibly duller (about 9 dB less energy at 4-8 kHz); this keeps the original export's sound.
• voices.bin, tokens.txt, lexicon-us-en.txt - from sherpa-onnx's kokoro-multi-lang-v1_0 release (Apache-2.0). Speaker ids: af_bella 2, af_heart 3, am_michael 16.
• espeak-ng-data/ - eSpeak NG phoneme data, GPL-3.0-or-later; source at that link.

pocket-2026-09

Pocket TTS by Kyutai, english_2026-09 (6 layers), licensed CC-BY-4.0. Exported to ONNX with KevinAHM/pocket-tts-onnx-export (Apache-2.0) using tools/english_2026-09.yaml, then:

• encoder.onnx has Pocket's learned BOS-before-voice embedding prepended to its output inside the graph (tools/bake_bos.py). Models from english_2026-04 on expect it and sherpa-onnx (written for english_2026-01) does not insert it; without it the output is noise.
• vocab.json, token_scores.json - the SentencePiece tokenizer converted with sherpa-onnx's scripts/pocket-tts/convert_tokenizer.py.
• lm_main, lm_flow and decoder are INT8 (dynamic quantization); encoder and text_conditioner are fp32.
• voices/ - the reference recordings each voice is cloned from, resampled to 24 kHz mono:
  • eponine.wav - VCTK speaker p262, sentence 23, enhanced by Kyutai (kyutai/tts-voices). From the CSTR VCTK Corpus, University of Edinburgh, CC-BY-4.0.
  • peter_yearsley.wav - from Voice-Zero, LibriVox recordings, CC0.

Both are the recordings behind Pocket TTS's own preset voices of those names.

Prohibited use (from Kyutai)

Use of our model must comply with all applicable laws and regulations and must not result in, involve, or facilitate any illegal, harmful, deceptive, fraudulent, or unauthorized activity. Prohibited uses include, without limitation, voice impersonation or cloning without explicit and lawful consent; misinformation, disinformation, or deception (including fake news, fraudulent calls, or presenting generated content as genuine recordings of real people or events); and the generation of unlawful, harmful, libelous, abusive, harassing, discriminatory, hateful, or privacy-invasive content.

Checks

Every pack was rendered through sherpa-onnx and transcribed with Whisper (base.en); word error rates matched the reference runtimes (kokoro-onnx, pocket-tts) on the same sentences, and the Kokoro output matched the original export's spectrum.