Lectus voices

Voice model packs for the Lectus offline book reader (iOS / Android). Files are downloaded by the app on demand; nothing here is meant to be used standalone, but feel free.

ru/ - Russian premium (Vosk-TTS 0.9 multi)

Derived from vosk-model-tts-ru-0.9-multi by Alpha Cephei (Apache 2.0). 5 speakers (3 female, 2 male), 22050 Hz, multistream VITS conditioned on a BERT prosody encoder.

file size what it is
ru/model.onnx 171 MB VITS synthesizer, unchanged upstream weights
ru/bert.int8.onnx 156 MB BERT prosody encoder, dynamic int8 quantization of the upstream fp32 model (624 MB -> 156 MB, no audible quality loss)
ru/dict.tsv 92 MB upstream pronunciation dictionary (2,037,867 words) converted to a sorted word<TAB>phonemes TSV for on-device mmap binary search; highest-probability variant kept per word
ru/vocab.txt 1.8 MB WordPiece vocab for the BERT tokenizer, unchanged

Changes vs upstream: int8 quantization of the BERT encoder and a mechanical dictionary format conversion. Model weights and voices are otherwise untouched.

kokoro/ - English, French, Spanish premium (Kokoro-82M v1.0)

Unmodified subset of kokoro-multi-lang-v1_0 packaged by k2-fsa from hexgrad/Kokoro-82M (Apache 2.0). 53 speakers, 24000 Hz; the app uses the English (US/GB), French (ff_siwis) and Spanish (ef_dora, em_alex) voices via sherpa-onnx. Chinese lexicons/FSTs from the original package are omitted; espeak-ng-data ships inside the app bundle instead.

file size
kokoro/model.onnx 326 MB
kokoro/voices.bin 28 MB
kokoro/lexicon-us-en.txt + lexicon-gb-en.txt 12 MB
kokoro/tokens.txt 1 KB

fp32 deliberately: dynamic int8 cuts size to 114 MB but ConvInteger makes CPU inference ~4x slower (RTF 0.66 vs 0.16) with audible quality loss.

uk/ - Ukrainian premium (StyleTTS2 ukrainian)

Derived from patriotyk/styletts2_ukrainian_multispeaker (MIT) and the styletts2-ukrainian space. 31 speakers, 24000 Hz, deterministic StyleTTS2 (no diffusion at inference).

file size what it is
uk/model.onnx 330 MB ONNX export of the multispeaker model (tokens + speed + style -> wav), exact parity with PyTorch
uk/stress.tsv 71 MB word-stress dictionary: package trie re-disambiguated by corpus-oracle majority vote (brown-uk corpus) + normative overrides; sorted word<TAB>positions for mmap binary search
uk/voices.bin 31 KB 31 speaker style vectors, [31, 256] float32 LE, order matches voices.json
uk/voices.json 1 KB speaker display names
uk/vocab.json 308 B tokenizer character vocabulary (184 symbols)
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support