Publish stable Russian Style-Bert v3/1000
Browse filesAdds the selected RU safetensors checkpoint, original five-style vectors, model config, training and selection audits, checksums, and verified inference instructions.
- Models/Russian_StyleBert_Pilot/v3-1000/README.md +47 -0
- Models/Russian_StyleBert_Pilot/v3-1000/SHA256SUMS +6 -0
- Models/Russian_StyleBert_Pilot/v3-1000/config.json +112 -0
- Models/Russian_StyleBert_Pilot/v3-1000/latest.safetensors +3 -0
- Models/Russian_StyleBert_Pilot/v3-1000/selection_audit.json +37 -0
- Models/Russian_StyleBert_Pilot/v3-1000/style_vectors.npy +3 -0
- Models/Russian_StyleBert_Pilot/v3-1000/training_audit.json +1 -0
Models/Russian_StyleBert_Pilot/v3-1000/README.md
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Russian GLaDOS Style-Bert-VITS2 — v3/1000
|
| 2 |
+
|
| 3 |
+
This is the selected stable Russian inference checkpoint for the GLaDOS voice.
|
| 4 |
+
It is a native Russian Style-Bert-VITS2 adaptation; English remains routed to
|
| 5 |
+
the unmodified original GLaDOS model.
|
| 6 |
+
|
| 7 |
+
## Files
|
| 8 |
+
|
| 9 |
+
- `latest.safetensors` — complete Russian acoustic checkpoint.
|
| 10 |
+
- `config.json` — 44.1 kHz model configuration and five-style mapping.
|
| 11 |
+
- `style_vectors.npy` — the five original GLaDOS style vectors.
|
| 12 |
+
- `training_audit.json` — immutable checkpoint/export provenance.
|
| 13 |
+
- `selection_audit.json` — held-out selection result and runtime contract.
|
| 14 |
+
- `SHA256SUMS` — hashes for every inference asset and audit.
|
| 15 |
+
|
| 16 |
+
## Styles
|
| 17 |
+
|
| 18 |
+
`Neutral`, `Standard`, `Deep`, `Light`, and `Standard_02` are preserved with
|
| 19 |
+
the original IDs 0 through 4.
|
| 20 |
+
|
| 21 |
+
## Download
|
| 22 |
+
|
| 23 |
+
```bash
|
| 24 |
+
hf download Random118/GLaDOS_TTS \
|
| 25 |
+
--include 'Models/Russian_StyleBert_Pilot/v3-1000/*' \
|
| 26 |
+
--local-dir ./GLaDOS_TTS
|
| 27 |
+
```
|
| 28 |
+
|
| 29 |
+
## Verified inference on the training host
|
| 30 |
+
|
| 31 |
+
The Russian frontend and hash checks are implemented in `orange-RAG` commit
|
| 32 |
+
`22c75c5`. From `/home/random/orange-RAG`:
|
| 33 |
+
|
| 34 |
+
```bash
|
| 35 |
+
GLADOS_ESPEAK_NG=/home/random/glados-tts/espeak-ng/bin/espeak-ng \
|
| 36 |
+
/home/random/glados-tts/.venv/bin/python \
|
| 37 |
+
-m glados_ru.run_verified_russian_glados \
|
| 38 |
+
--profile v3-1000 \
|
| 39 |
+
--text 'Закро́йте две́рь и сохрани́те па́роль.' \
|
| 40 |
+
--style Neutral \
|
| 41 |
+
--output /tmp/glados-ru.wav
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
The default frontend applies `RuAccent` to plain Russian. Explicit combining
|
| 45 |
+
acute accents (`о́`) and `ё` take priority. The output is mono PCM16 at 44.1
|
| 46 |
+
kHz. The verified one-shot CPU process used about 1.65 GiB RAM and 0 MiB VRAM.
|
| 47 |
+
|
Models/Russian_StyleBert_Pilot/v3-1000/SHA256SUMS
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dd8c8987ddd003236589863b3d483d0f50f27e9ea17c92ed0bf839a4818a53ae latest.safetensors
|
| 2 |
+
d8baa2c240562034431f6f5ebf2d107d2beb21768608c30348b945482138ff30 config.json
|
| 3 |
+
80cc6ff2333eea2644b5008bf4fce1e838038b8431d72adb8c282aa29c2e274e style_vectors.npy
|
| 4 |
+
59ac5781a47e6409b61f1b9f9d23695a1005f2fedc7b7064b6259310e34766d1 training_audit.json
|
| 5 |
+
0c368b610f2333f650c4a17f357590fb22d43aeeab4a7b7588adcf80261d2346 selection_audit.json
|
| 6 |
+
d67023129306ba3e2eb9bbf7b0e70111f98ede4a2f307935afb7685117346383 README.md
|
Models/Russian_StyleBert_Pilot/v3-1000/config.json
ADDED
|
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "Portal_GLaDOS_v1",
|
| 3 |
+
"train": {
|
| 4 |
+
"log_interval": 200,
|
| 5 |
+
"eval_interval": 1000,
|
| 6 |
+
"seed": 42,
|
| 7 |
+
"epochs": 1000,
|
| 8 |
+
"learning_rate": 0.0002,
|
| 9 |
+
"betas": [
|
| 10 |
+
0.8,
|
| 11 |
+
0.99
|
| 12 |
+
],
|
| 13 |
+
"eps": 1e-09,
|
| 14 |
+
"batch_size": 4,
|
| 15 |
+
"bf16_run": false,
|
| 16 |
+
"lr_decay": 0.99995,
|
| 17 |
+
"segment_size": 16384,
|
| 18 |
+
"init_lr_ratio": 1,
|
| 19 |
+
"warmup_epochs": 0,
|
| 20 |
+
"c_mel": 45,
|
| 21 |
+
"c_kl": 1.0,
|
| 22 |
+
"skip_optimizer": false,
|
| 23 |
+
"freeze_ZH_bert": false,
|
| 24 |
+
"freeze_JP_bert": false,
|
| 25 |
+
"freeze_EN_bert": false,
|
| 26 |
+
"freeze_style": false,
|
| 27 |
+
"freeze_encoder": false,
|
| 28 |
+
"freeze_decoder": false
|
| 29 |
+
},
|
| 30 |
+
"data": {
|
| 31 |
+
"use_jp_extra": false,
|
| 32 |
+
"training_files": "Data\\Portal_GLaDOS_v1\\train.list",
|
| 33 |
+
"validation_files": "Data\\Portal_GLaDOS_v1\\val.list",
|
| 34 |
+
"max_wav_value": 32768.0,
|
| 35 |
+
"sampling_rate": 44100,
|
| 36 |
+
"filter_length": 2048,
|
| 37 |
+
"hop_length": 512,
|
| 38 |
+
"win_length": 2048,
|
| 39 |
+
"n_mel_channels": 128,
|
| 40 |
+
"mel_fmin": 0.0,
|
| 41 |
+
"mel_fmax": null,
|
| 42 |
+
"add_blank": true,
|
| 43 |
+
"n_speakers": 1,
|
| 44 |
+
"cleaned_text": true,
|
| 45 |
+
"num_styles": 5,
|
| 46 |
+
"style2id": {
|
| 47 |
+
"Neutral": 0,
|
| 48 |
+
"Standard": 1,
|
| 49 |
+
"Deep": 2,
|
| 50 |
+
"Light": 3,
|
| 51 |
+
"Standard_02": 4
|
| 52 |
+
},
|
| 53 |
+
"spk2id": {
|
| 54 |
+
"Portal_GLaDOS_v1": 0
|
| 55 |
+
}
|
| 56 |
+
},
|
| 57 |
+
"model": {
|
| 58 |
+
"use_spk_conditioned_encoder": true,
|
| 59 |
+
"use_noise_scaled_mas": true,
|
| 60 |
+
"use_mel_posterior_encoder": false,
|
| 61 |
+
"use_duration_discriminator": true,
|
| 62 |
+
"inter_channels": 192,
|
| 63 |
+
"hidden_channels": 192,
|
| 64 |
+
"filter_channels": 768,
|
| 65 |
+
"n_heads": 2,
|
| 66 |
+
"n_layers": 6,
|
| 67 |
+
"kernel_size": 3,
|
| 68 |
+
"p_dropout": 0.1,
|
| 69 |
+
"resblock": "1",
|
| 70 |
+
"resblock_kernel_sizes": [
|
| 71 |
+
3,
|
| 72 |
+
7,
|
| 73 |
+
11
|
| 74 |
+
],
|
| 75 |
+
"resblock_dilation_sizes": [
|
| 76 |
+
[
|
| 77 |
+
1,
|
| 78 |
+
3,
|
| 79 |
+
5
|
| 80 |
+
],
|
| 81 |
+
[
|
| 82 |
+
1,
|
| 83 |
+
3,
|
| 84 |
+
5
|
| 85 |
+
],
|
| 86 |
+
[
|
| 87 |
+
1,
|
| 88 |
+
3,
|
| 89 |
+
5
|
| 90 |
+
]
|
| 91 |
+
],
|
| 92 |
+
"upsample_rates": [
|
| 93 |
+
8,
|
| 94 |
+
8,
|
| 95 |
+
2,
|
| 96 |
+
2,
|
| 97 |
+
2
|
| 98 |
+
],
|
| 99 |
+
"upsample_initial_channel": 512,
|
| 100 |
+
"upsample_kernel_sizes": [
|
| 101 |
+
16,
|
| 102 |
+
16,
|
| 103 |
+
8,
|
| 104 |
+
2,
|
| 105 |
+
2
|
| 106 |
+
],
|
| 107 |
+
"n_layers_q": 3,
|
| 108 |
+
"use_spectral_norm": false,
|
| 109 |
+
"gin_channels": 256
|
| 110 |
+
},
|
| 111 |
+
"version": "2.4.1"
|
| 112 |
+
}
|
Models/Russian_StyleBert_Pilot/v3-1000/latest.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dd8c8987ddd003236589863b3d483d0f50f27e9ea17c92ed0bf839a4818a53ae
|
| 3 |
+
size 198808124
|
Models/Russian_StyleBert_Pilot/v3-1000/selection_audit.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"status": "selected_stable_default",
|
| 3 |
+
"profile": "v3-1000",
|
| 4 |
+
"checkpoint_sha256": "dd8c8987ddd003236589863b3d483d0f50f27e9ea17c92ed0bf839a4818a53ae",
|
| 5 |
+
"config_sha256": "d8baa2c240562034431f6f5ebf2d107d2beb21768608c30348b945482138ff30",
|
| 6 |
+
"style_vectors_sha256": "80cc6ff2333eea2644b5008bf4fce1e838038b8431d72adb8c282aa29c2e274e",
|
| 7 |
+
"training_audit_sha256": "59ac5781a47e6409b61f1b9f9d23695a1005f2fedc7b7064b6259310e34766d1",
|
| 8 |
+
"styles": {
|
| 9 |
+
"Neutral": 0,
|
| 10 |
+
"Standard": 1,
|
| 11 |
+
"Deep": 2,
|
| 12 |
+
"Light": 3,
|
| 13 |
+
"Standard_02": 4
|
| 14 |
+
},
|
| 15 |
+
"held_out_russian_100": {
|
| 16 |
+
"gigaam": {
|
| 17 |
+
"exact": 37,
|
| 18 |
+
"wer": 0.1968
|
| 19 |
+
},
|
| 20 |
+
"whisper": {
|
| 21 |
+
"exact": 47,
|
| 22 |
+
"wer": 0.1604
|
| 23 |
+
},
|
| 24 |
+
"both_exact": 26
|
| 25 |
+
},
|
| 26 |
+
"runtime": {
|
| 27 |
+
"language": "RU",
|
| 28 |
+
"sample_rate": 44100,
|
| 29 |
+
"output": "mono PCM16 WAV",
|
| 30 |
+
"automatic_stress": "RuAccent turbo3.1; explicit U+0301 and ё take priority",
|
| 31 |
+
"verified_peak_rss_kib": 1685160,
|
| 32 |
+
"expected_vram_mib": 0,
|
| 33 |
+
"frontend_commit": "22c75c5"
|
| 34 |
+
},
|
| 35 |
+
"english_route": "Use the separate, unmodified original GLaDOS checkpoint for EN.",
|
| 36 |
+
"excluded_candidate": "The later Baya ending-correction checkpoint is still pending human listening and is not this release."
|
| 37 |
+
}
|
Models/Russian_StyleBert_Pilot/v3-1000/style_vectors.npy
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:80cc6ff2333eea2644b5008bf4fce1e838038b8431d72adb8c282aa29c2e274e
|
| 3 |
+
size 5248
|
Models/Russian_StyleBert_Pilot/v3-1000/training_audit.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"status": "trained", "checkpoint_sha256": "dd8c8987ddd003236589863b3d483d0f50f27e9ea17c92ed0bf839a4818a53ae", "step": 1000, "source_resume": "/home/random/glados-tts/native-ru/pilot-run-v3/resume-step1000.pt", "source_resume_sha256": "ddccfa01045d7906a356e6e3cec81aa80299e409c062c699b9c5334681e1c8d2"}
|