Pocket TTS Russian (Shtorm) for Core ML

ArtShtorm's Shtorm-PocketTTS-RU, a Russian model with Kyutai's Pocket TTS architecture, exported to Core ML for TextSniper by MacEpic. TextSniper downloads it when someone chooses Pocket TTS for Russian in its settings.

Model

Weights: model_fast.safetensors of ArtShtorm/Shtorm-PocketTTS-RU v2 (revision 33f6c328), the fast variant, by ArtShtorm, a community model with Kyutai's Pocket TTS architecture, licensed CC BY 4.0. Unchanged apart from the conversion.
Tokenizer: its BPE SentencePiece model, converted to tokenizer.json.
It reads clearly only with stress marks: TextSniper marks the text with RUAccent first (MacEpic/ruaccent-coreml).
Converted with the export_coreml example of xn-ptts 0.6.0 (Apache 2.0 / MIT), which also reads it: bundle.json lists every file with its size and SHA-256 (15 files, 398,732,561 bytes).

Voices

Each voice in voices/ was made with Shtorm's own voice encoder from the recording below, trimmed of silence, with the model's bos_before_voice prepended (emb, xn-ptts's voice format).

Voice	Recording	License
estelle	`unmute-prod-website/developpeuse-3.wav` in kyutai/tts-voices (revision `1fc7395b`), Kyutai's own recording	CC0
alba	`alba-mackenna/casual.wav` in kyutai/tts-voices, voice-acted by Alba MacKenna	CC BY 4.0
anna	`vctk/p228_023_enhanced.wav` in kyutai/tts-voices, speaker p228 of the VCTK corpus	CC BY 4.0
bill	`voice-zero/bill_boerst.wav` in kyutai/tts-voices, from Voice-Zero (LibriVox)	CC0

License

CC BY 4.0, as the model: credit ArtShtorm's Shtorm-PocketTTS-RU, Kyutai's Pocket TTS and the voices' sources above.

Downloads last month: -

Model tree for MacEpic/pocket-tts-ru-coreml

Base model

kyutai/pocket-tts

Finetuned

ArtShtorm/Shtorm-PocketTTS-RU

Quantized

(1)

this model