ModelListedExternal

Subly Ru Tts

A fully on-device Russian text-to-speech voice for Apple platforms (iOS/macOS), plus everything needed to turn raw Russian text into the phoneme stream the model expects. The acoustic model is a StyleTTS2 fine-tune...

Sourcehuggingface.co/nakash1987/subly-ru-tts/tree/7ec4d04796dccaf448729c7fa1c55cf1a2d3a556

KindOtherVersionRevision 7ec4d04

Model source

Kind: Other
File size: 243 B–128 MB
Version: vRevision 7ec4d04
Updated: Sep 13, 2026
Downloads: 0
Source: Hugging Face

Source description

Subly RU TTS — On-device Russian voice "Natasha" (StyleTTS2 → Core ML)

A fully on-device Russian text-to-speech voice for Apple platforms (iOS/macOS), plus everything needed to turn raw Russian text into the phoneme stream the model expects. The acoustic model is a StyleTTS2 fine-tune exported to Core ML, designed to run through the FluidAudio StyleTTS2 runtime. No server, no network at inference time, no Python.

Built for the Subly app, released openly so anyone can use it. 24 kHz, female voice, neutral narration style.

What's in this repo

Path	What it is	Size
`models/*.mlmodelc`	14 StyleTTS2 stages compiled to Core ML (FLOAT32): `text_encoder`, `bert` (+`t64/t128/t256` token buckets), `ref_encoder`, `fused_diffusion_sampler` (+buckets), `duration_predictor`, `fused_f0n_har_source`, `decoder_pre`, `decoder_upsample`	~867 MB
`ru_stress.bin`	Word→stress dictionary, 3.19 M entries (mmap, binary-searchable)	87 MB
`ru_accent_oov_fp32.mlmodelc`	Char-level RoFormer that predicts stress for out-of-dictionary words	2.2 MB
`ru_yo.bin`	ё-restoration dictionary, 85.5 K words (е→ё)	2.5 MB
`ru_natasha_ref.wav`	Reference clip — StyleTTS2 timbre is zero-shot from this audio (24 kHz mono)	156 KB
`ru_g2p.py`	Reference grapheme→IPA phonemizer (rule-based Russian phonology)	—
`manifest.json`	Flat list of every file + size (for on-demand downloaders)	—

The pipeline

Russian text becomes audio in four steps. The first three are CPU-cheap; only the last touches the neural models.

raw text
  → ё-restoration         (ru_yo.bin)          "елка" → "ёлка"
  → stress marking        (ru_stress.bin,      "ёлка" → "+ёлка"   (+ before the stressed vowel)
                           OOV model for misses)
  → grapheme→IPA          (ru_g2p.py)          "+ёлка" → "jˈolkə"
  → StyleTTS2 synthesis   (models/, ref clip)  IPA → 24 kHz waveform

This mirrors the training pipeline (RUAccent stress + a rule-based G2P), so the IPA fed at inference matches what the model was trained on.

Recommended runtime parameters

Param	Value	Note
`alpha`	`0.5`	reference/diffusion style blend
`beta`	`0.9`	more diffusion-driven prosody = livelier
`noiseSeed`	`1`	a pleasant default; only changes the per-utterance sample
compute units	CPU-only	the diffusion stage is FP32 (can't run on ANE); under MPS the ref-encoder mixes FP16/FP32 and crashes — run on CPU

Using it with FluidAudio (Swift)

import FluidAudio

// Point the manager at the `models/` directory from this repo.
let manager = StyleTTS2Manager(directory: modelsDir, computeUnits: .cpuOnly)
try await manager.initialize()

let ipa = "prjɪvjˈɛt ! ˈɛtə tjˈɛst rˈusskəvə djˈiktərə ."   // from the pipeline above
let samples = try await manager.synthesize(
    ipa: ipa,
    referenceAudioURL: refClipURL,   // ru_natasha_ref.wav
    alpha: 0.5, beta: 0.9, noiseSeed: 1
)   // -> [Float] PCM @ 24 kHz

The IPA is tokenized by FluidAudio's StyleTTS2TextCleaner (the StyleTTS2 178-symbol set); any symbol outside that set is dropped, so feed clean IPA.

File formats (language-agnostic)

So you can read these from any language, not just Swift.

`ru_stress.bin` and `ru_yo.bin`

Little-endian, sorted by UTF-8 byte order (so a byte-wise binary search works):

ru_stress.bin:  ["RUS1"][u32 count][offsets: count × u32][records]
  record = [u8 keyLen][key UTF-8][u8 pos]
  pos = char index of the stressed vowel (where '+' is inserted); 255 = unstressed.

ru_yo.bin:      ["RUYO"][u32 count][offsets: count × u32][records]
  record = [u8 keyLen][key UTF-8][u8 posCount][posCount × u8 charIndex]
  charIndex = positions in the (lowercased) word where 'е' should become 'ё'.

offsets[i] is the byte offset of record i relative to the start of the records section (= 8 + count*4). Binary-search by reading offsets[mid], then the key.

`ru_accent_oov_fp32.mlmodelc` (OOV stress model)

Char-level token classifier (converted from RUAccent's nn_accent).

Input input_ids: [1, 48] int32 — [bos]=2, then chars by the vocab below, [eos]=3, padded with [pad]=0.
Output logits: [1, 48, 3] — labels 0=NO, 1=STRESS_PRIMARY, 2=STRESS_SECONDARY.
Softmax per position; mark a + before char i-1 when position i is argmax PRIMARY with probability ≥ 0.55.

Vocab (id → token): 0 [pad], 1 [unk], 2 [bos], 3 [eos], 4 ', 5 -, 6 ., 7 ?, 8 `, 9 c, 10 e, 11 ́ (U+0301), then Cyrillic 12 а ... 43 я, 44 ё.

Model details

Architecture: StyleTTS2 (LibriTTS config, HiFi-GAN decoder), multilingual PL-BERT (ru), 80-mel / n_fft 2048 / win 1200 / hop 300, 24 kHz.
Voice / training data: SOVA Dataset — TTS Natasha (~12.4 h, one female speaker).
Phonemization: stress from RUAccent (dictionary + char model + ё-restoration), then a rule-based G2P covering palatalization, two-degree vowel reduction, iotation, final devoicing and regressive voicing assimilation, plus the г→в genitive exception.
Core ML export: all stages FLOAT32 (the diffusion sampler overflows in FP16; other stages were kept FP32 for fidelity). Token-axis buckets 57/64/128/256.

Limitations

Homographs (за́мок / замо́к, по́лет / полёт) take the dictionary's default stress — there is no context disambiguator here (that model is ~350 MB and was left out). They're rare in narration.
Numbers/currency are not spelled out — normalize digits to words upstream (a bare 123 won't be voiced).
Size: FP32 models are ~867 MB. For shipping, an FP16 re-export of the non-diffusion stages cuts this substantially, but re-validate quality by ear.
CPU-only at inference (see above).

License & attribution

This repository is released under CC BY 4.0 — use it commercially or non-commercially, just keep the attributions below.

Voice / reference audio derive from the SOVA Dataset (TTS Natasha), © Virtual Assistant, LLC, licensed CC BY 4.0.
StyleTTS2 — MIT (Aaron (Yinghao) Li et al.).
multilingual PL-BERT — MIT (Papercup).
RUAccent (stress dictionary, OOV model, ё-dictionary are derived from it) — Apache-2.0 (Den4ikAI).

Please credit "SOVA TTS Natasha (CC BY 4.0, Virtual Assistant LLC)" and this repo when you use the voice.

Sources

1 source

Hugging FacePrimaryVersion 7ec4d04796dccaf448729c7fa1c55cf1a2d3a556

huggingface.co/nakash1987/subly-ru-tts/tree/7ec4d04796dccaf448729c7fa1c55cf1a2d3a556

Verified Sep 13

Model artifacts

64 artifacts

models/bert_fp16_t128.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 8ccaff5aae43…c14a · Hugging Face

Download

models/bert_fp16_t128.mlmodelc/coremldata.bin

bin · 441 B · SHA-256 1b249e5af898…996f · Hugging Face

Download

models/bert_fp16_t128.mlmodelc/weights/weight.bin

bin · 23.1 MB · SHA-256 76d33b62ef8b…606e · Hugging Face

Download

models/bert_fp16_t256.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 aab6f4e80f3a…a9e3 · Hugging Face

Download

models/bert_fp16_t256.mlmodelc/coremldata.bin

bin · 441 B · SHA-256 a6be4cb0edd0…07f1 · Hugging Face

Download

models/bert_fp16_t256.mlmodelc/weights/weight.bin

bin · 23.3 MB · SHA-256 40c3e10b5afa…a15d · Hugging Face

Download

models/bert_fp16_t64.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 f7beefe8f597…f3d2 · Hugging Face

Download

models/bert_fp16_t64.mlmodelc/coremldata.bin

bin · 437 B · SHA-256 fb93a6c5cc45…10f8 · Hugging Face

Download

models/bert_fp16_t64.mlmodelc/weights/weight.bin

bin · 23.1 MB · SHA-256 bbd7fef72cd0…4cdd · Hugging Face

Download

models/bert_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 1dc0b35fffb6…345a · Hugging Face

Download

models/bert_fp16.mlmodelc/coremldata.bin

bin · 437 B · SHA-256 928469517e42…d90e · Hugging Face

Download

models/bert_fp16.mlmodelc/weights/weight.bin

bin · 23.1 MB · SHA-256 5d98fa393bda…5fd5 · Hugging Face

Download

models/decoder_pre_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 9376a753a257…31a0 · Hugging Face

Download

models/decoder_pre_fp16.mlmodelc/coremldata.bin

bin · 490 B · SHA-256 b1dce025e38b…9324 · Hugging Face

Download

models/decoder_pre_fp16.mlmodelc/weights/weight.bin

bin · 128 MB · SHA-256 c2427eac607c…ea80 · Hugging Face

Download

models/decoder_upsample_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 a71b19d581e7…c0a4 · Hugging Face

Download

models/decoder_upsample_fp16.mlmodelc/coremldata.bin

bin · 467 B · SHA-256 161c40ceadc9…4684 · Hugging Face

Download

models/decoder_upsample_fp16.mlmodelc/weights/weight.bin

bin · 78.9 MB · SHA-256 b3b7c42b28c2…937f · Hugging Face

Download

models/duration_predictor_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 7c6963b51287…4eac · Hugging Face

Download

models/duration_predictor_fp16.mlmodelc/coremldata.bin

bin · 470 B · SHA-256 829e39feda9e…1798 · Hugging Face

Download

models/duration_predictor_fp16.mlmodelc/weights/weight.bin

bin · 29.6 MB · SHA-256 2f5face6829b…97a5 · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t128.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 adbb9416c3b4…af7d · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t128.mlmodelc/coremldata.bin

bin · 462 B · SHA-256 073af3e2b1da…7fe1 · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t128.mlmodelc/weights/weight.bin

bin · 94.1 MB · SHA-256 38fa4543513f…37c9 · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t256.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 6de3456a7b67…f50c · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t256.mlmodelc/coremldata.bin

bin · 462 B · SHA-256 8541ab25bb7a…b517 · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t256.mlmodelc/weights/weight.bin

bin · 94.1 MB · SHA-256 38fa4543513f…37c9 · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t64.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 90fb7f8121c1…fc9b · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t64.mlmodelc/coremldata.bin

bin · 461 B · SHA-256 07d596ba5d69…52f5 · Hugging Face

Download

models/fused_diffusion_sampler_fp16_t64.mlmodelc/weights/weight.bin

bin · 94.1 MB · SHA-256 38fa4543513f…37c9 · Hugging Face

Download

models/fused_diffusion_sampler_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 790c55beced8…6f0e · Hugging Face

Download

models/fused_diffusion_sampler_fp16.mlmodelc/coremldata.bin

bin · 461 B · SHA-256 ebddfa2bd931…30f2 · Hugging Face

Download

models/fused_diffusion_sampler_fp16.mlmodelc/weights/weight.bin

bin · 94.1 MB · SHA-256 38fa4543513f…37c9 · Hugging Face

Download

models/fused_f0n_har_source.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 2a942bab9c30…53f2 · Hugging Face

Download

models/fused_f0n_har_source.mlmodelc/coremldata.bin

bin · 441 B · SHA-256 777f486931a7…4328 · Hugging Face

Download

models/fused_f0n_har_source.mlmodelc/weights/weight.bin

bin · 32.1 MB · SHA-256 0103bde900eb…b6ff · Hugging Face

Download

models/ref_encoder_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 6644754aa623…f644 · Hugging Face

Download

models/ref_encoder_fp16.mlmodelc/coremldata.bin

bin · 372 B · SHA-256 898ccfedf04c…6030 · Hugging Face

Download

models/ref_encoder_fp16.mlmodelc/weights/weight.bin

bin · 106 MB · SHA-256 5dc31c37b785…6995 · Hugging Face

Download

models/text_encoder_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 632216b5d62f…0150 · Hugging Face

Download

models/text_encoder_fp16.mlmodelc/coremldata.bin

bin · 452 B · SHA-256 559d31acf61a…4d73 · Hugging Face

Download

models/text_encoder_fp16.mlmodelc/weights/weight.bin

bin · 21.4 MB · SHA-256 c2596ba452e8…7f54 · Hugging Face

Download

ru_accent_oov_fp32.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 32ab2cfacbed…c73e · Hugging Face

Download

ru_accent_oov_fp32.mlmodelc/coremldata.bin

bin · 374 B · SHA-256 b4a962493dff…5cad · Hugging Face

Download

ru_accent_oov_fp32.mlmodelc/weights/weight.bin

bin · 2.08 MB · SHA-256 4bfdcddea36c…7730 · Hugging Face

Download

ru_stress.bin

bin · 86.8 MB · SHA-256 2534986fe0f4…226c · Hugging Face

Download

ru_yo.bin

bin · 2.39 MB · SHA-256 7f5d4dd9202e…12d2 · Hugging Face

Download

styletts2/bert_fp16_t128.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 8ccaff5aae43…c14a · Hugging Face

Download

styletts2/bert_fp16_t128.mlmodelc/coremldata.bin

bin · 441 B · SHA-256 1b249e5af898…996f · Hugging Face

Download

styletts2/bert_fp16_t128.mlmodelc/weights/weight.bin

bin · 23.1 MB · SHA-256 76d33b62ef8b…606e · Hugging Face

Download

styletts2/bert_fp16_t256.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 aab6f4e80f3a…a9e3 · Hugging Face

Download

styletts2/bert_fp16_t256.mlmodelc/coremldata.bin

bin · 441 B · SHA-256 a6be4cb0edd0…07f1 · Hugging Face

Download

styletts2/bert_fp16_t256.mlmodelc/weights/weight.bin

bin · 23.3 MB · SHA-256 40c3e10b5afa…a15d · Hugging Face

Download

styletts2/bert_fp16_t64.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 f7beefe8f597…f3d2 · Hugging Face

Download

styletts2/bert_fp16_t64.mlmodelc/coremldata.bin

bin · 437 B · SHA-256 fb93a6c5cc45…10f8 · Hugging Face

Download

styletts2/bert_fp16_t64.mlmodelc/weights/weight.bin

bin · 23.1 MB · SHA-256 bbd7fef72cd0…4cdd · Hugging Face

Download

styletts2/bert_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 1dc0b35fffb6…345a · Hugging Face

Download

styletts2/bert_fp16.mlmodelc/coremldata.bin

bin · 437 B · SHA-256 928469517e42…d90e · Hugging Face

Download

styletts2/bert_fp16.mlmodelc/weights/weight.bin

bin · 23.1 MB · SHA-256 5d98fa393bda…5fd5 · Hugging Face

Download

styletts2/decoder_pre_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 9376a753a257…31a0 · Hugging Face

Download

styletts2/decoder_pre_fp16.mlmodelc/coremldata.bin

bin · 490 B · SHA-256 b1dce025e38b…9324 · Hugging Face

Download

styletts2/decoder_pre_fp16.mlmodelc/weights/weight.bin

bin · 128 MB · SHA-256 c2427eac607c…ea80 · Hugging Face

Download

styletts2/decoder_upsample_fp16.mlmodelc/analytics/coremldata.bin

bin · 243 B · SHA-256 a71b19d581e7…c0a4 · Hugging Face

Download

styletts2/decoder_upsample_fp16.mlmodelc/coremldata.bin

bin · 467 B · SHA-256 161c40ceadc9…4684 · Hugging Face

Download

Source excerpts

2 excerpts

ReadmeSep 13huggingface.co/nakash1987/subly-ru-tts

--- license: cc-by-4.0 language: - ru pipeline_tag: text-to-speech tags: - text-to-speech - tts - russian - styletts2 - coreml - on-device - ios - macos library_name: coreml --- # Subly RU TTS — On-device Russian voice "Natasha" (StyleTTS2 → Core ML) A **fully on-device Russian text-to-speech voice** for Apple platforms (iOS/macOS), plus everything needed to turn raw Russian text into the phoneme stream the model expects. The acoustic model is a [StyleTTS2](https://github.com/yl4579/StyleTTS2) fine-tune exported to **Core ML**, designed to run through the [FluidAudio](https://github.com/FluidInference/FluidAudio) `StyleTTS2` runtime. No server, no network at inference time, no Python. > Built for the **Subly** app, released openly so anyone can use it. 24 kHz, female > voice, neutral narration style. --- ## What's in this repo | Path | What it is | Size | |---|---|---| | `models/*.mlmodelc` | 14 StyleTTS2 stages compiled to Core ML (FLOAT32): `text_encoder`, `bert` (+`t64/t128/t256` token buckets), `ref_encoder`, `fused_diffusion_sampler` (+buckets), `duration_predictor`, `fused_f0n_har_source`, `decoder_pre`, `decoder_upsample` | ~867 MB | | `ru_stress.bin` | Word→stress dictionary, **3.19 M** entries (mmap, binary-searchable) | 87 MB | | `ru_accent_oov_fp32.mlmodelc` | Char-level RoFormer that predicts stress for **out-of-dictionary** words | 2.2 MB | | `ru_yo.bin` | ё-restoration dictionary, **85.5 K** words (е→ё) | 2.5 MB | | `ru_natasha_ref.wav` | Reference clip — StyleTTS2 timbre is zero-shot from this audio (24 kHz mono) | 156 KB | | `ru_g2p.py` | Reference grapheme→IPA phonemizer (rule-based Russian phonology) | — | | `manifest.json` | Flat list of every file + size (for on-demand downloaders) | — | --- ## The pipeline Russian text becomes audio in four steps. The first three are CPU-cheap; only the last touches the neural models. ``` raw text → ё-restoration (ru_yo.bin) "елка" → "ёлка" → stress marking (ru_stress.bin, "ёлка" → "+ёлка" (+ before the stressed vowel) OOV model for misses) → grapheme→IPA (ru_g2p.py) "+ёлка" → "jˈolkə" → StyleTTS2 synthesis (models/, ref clip) IPA → 24 kHz waveform ``` This mirrors the training pipeline (RUAccent stress + a rule-based G2P), so the IPA fed at inference matches what the model was trained on. ### Recommended runtime parameters | Param | Value | Note | |---|---|---| | `alpha` | `0.5` | r...

NoteSep 13huggingface.co/nakash1987/subly-ru-tts

Source context: 0 downloads · 1 likes · Pipeline text-to-speech · Library coreml · Repo nakash1987/subly-ru-tts

Subly Ru Tts

Subly RU TTS — On-device Russian voice "Natasha" (StyleTTS2 → Core ML)

What's in this repo

The pipeline

Recommended runtime parameters

Using it with FluidAudio (Swift)

File formats (language-agnostic)

ru_stress.bin and ru_yo.bin

ru_accent_oov_fp32.mlmodelc (OOV stress model)

Model details

Limitations

License & attribution

`ru_stress.bin` and `ru_yo.bin`

`ru_accent_oov_fp32.mlmodelc` (OOV stress model)