Add model-anchored 0234 voice comparison

Publishes a clearly marked A/B review against the official Style-Bert-VITS2 NeutralStyle output before regenerating the full Russian pack.

Models/Russian_CosyVoice3/model_anchor_review/0234/README.md

@@ -0,0 +1,19 @@
+# 0234 model-anchor review
+
+This review set compares the published Russian clip with two variants anchored to
+the official `NeutralStyle` output of `WarriorMama777/GLaDOS_TTS`.
+
+| File | Voice anchor | Median F0 | Anchor cosine | GigaAM WER |
+| --- | --- | ---: | ---: | ---: |
+| `00-stylebert-neutral-reference.wav` | Official model output | 155.49 Hz | 1.000000 | n/a |
+| `01-current-published.wav` | Per-line original game recording | 146.76 Hz | 0.685028 | 0.000000 |
+| `02-model-anchored-balanced.wav` | Official model output | 154.6 Hz | 0.845889 | 0.000000 |
+| `03-model-anchored-natural-f0.wav` | Official model output | 160.05 Hz | 0.845073 | 0.000000 |
+
+Both model-anchored candidates use the original game clip only for final duration
+and pacing. A short source pause after `разногласия` preserves the canonical text
+through voice conversion without changing the spoken words.
+
+`02-model-anchored-balanced.wav` is the automatic recommendation: its pitch is
+aligned to the official model output, its duration matches the original line, and
+GigaAM recognizes the complete Russian text exactly.

Models/Russian_CosyVoice3/model_anchor_review/0234/model-anchored-balanced.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9c47afdef454c0b61d853720501c77d4ee24e9c4ed52665dd9600dcd1c2a336
+size 555512

Models/Russian_CosyVoice3/model_anchor_review/0234/model-anchored-natural-f0.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:288b37da207bc8ea3478e598f81751ce3895ebd09721544318291d62adf76c2e
+size 555512

README.md

@@ -74,6 +74,37 @@ The Russian pack contains 20 WAV samples, a complete bilingual manifest, per-cli
 
 Live preview and contributor fork: [Random118/GLaDOS_TTS](https://huggingface.co/Random118/GLaDOS_TTS)
 
+> **Russian voice/style revision in progress.** The first public pack used per-line
+> game recordings as speaker prompts. The replacement pipeline anchors speaker
+> identity and median F0 to this repository's official Style-Bert-VITS2
+> `NeutralStyle` output. The upstream PR is paused until this A/B review is accepted.
+
+#### Model-anchored 0234 A/B review
+
+**Official model voice — Style-Bert-VITS2 `NeutralStyle` reference**
+
+<audio controls preload="none" src="https://github.com/WarriorMama777/imgup/raw/main/img/__Repository/huggingface/AI/GLaDOS_TTS/Portal_GLaDOS_SBV2_v1_neutral_original_en_short_comp.mp3"></audio>
+
+**Current published Russian baseline — game-recording prompt, voice score 0.685**
+
+<audio controls preload="none" src="https://huggingface.co/Random118/GLaDOS_TTS/resolve/main/Models/Russian_CosyVoice3/samples/0234.wav"></audio>
+
+**Recommended model-anchored candidate — model F0, voice score 0.846, exact ASR**
+
+<audio controls preload="none" src="https://huggingface.co/Random118/GLaDOS_TTS/resolve/main/Models/Russian_CosyVoice3/model_anchor_review/0234/model-anchored-balanced.wav"></audio>
+
+**Model-anchored candidate — natural VC F0, voice score 0.845, exact ASR**
+
+<audio controls preload="none" src="https://huggingface.co/Random118/GLaDOS_TTS/resolve/main/Models/Russian_CosyVoice3/model_anchor_review/0234/model-anchored-natural-f0.wav"></audio>
+
+| Version | Median F0 | Cosine similarity to model voice | GigaAM WER |
+| --- | ---: | ---: | ---: |
+| Current published baseline | 146.76 Hz | 0.685 | 0.000 |
+| Model-anchored, model F0 | 154.6 Hz | **0.846** | **0.000** |
+| Model-anchored, natural F0 | 160.05 Hz | 0.845 | **0.000** |
+
+[Detailed 0234 comparison notes](https://huggingface.co/Random118/GLaDOS_TTS/blob/main/Models/Russian_CosyVoice3/model_anchor_review/0234/README.md)
+
 #### Russian audio examples
 
 **0000 —** «И снова здравствуйте. Добро пожаловать в компьютеризированный центр обогащения „Апертур Наука“».