coriollon/whisper-large-v3-turbo-russian

Whisper-large-v3-turbo Russian (Stage AW — long-form)

A Russian-specialized openai/whisper-large-v3-turbo fine-tune with English preserved. This release (Stage AW) continues Stage AS v2 with a targeted second pass that fixes long-form punctuation collapse — the model's tendency to stop emitting ., ,, ?, ! in the second half of 20-30 s meeting-style segments.

Aggregate WER / Cap F1 / Sentence F1 stay at Stage AS v2 levels; punctuation density in long turns grows ~3× and n_valid rises from 53/150 to 140/150.

Headline metrics

General ASR (6 RU test sets, N=500 each, augmented refs):

Metric	Stage AS v2 (prev)	Stage AW (this)	Δ
WER (aggregate)	9.02 %	9.07 %	+0.05 pp
PER F1 (punct macro F1)	0.398	0.388	−0.010
Cap F1 (word-level case-match)	0.822	0.821	−0.001
Sent F1 (sentence-boundary F1)	0.811	0.818	+0.007

Long-form punct density (150 held-out concat-turns, 12-29 s each):

Metric	Stage AS v2	Stage AW	Δ
first-half density mean	0.074	0.214	×2.9
second-half density mean	0.090	0.238	×2.7
ratio_med (d2/d1)	1.14	1.00	parity
n_valid (both halves punctuated)	53/150	140/150	×2.6

Per-dataset WER / PER F1

Dataset	AS v2 WER	AW WER	AS v2 PER	AW PER
Common Voice 21 RU	5.01 %	4.96 %	0.387	0.393
RuLibriSpeech	8.05 %	7.41 %	0.278	0.335
Sberdevices Golos far-field	9.42 %	9.42 %	0.596	0.487
Sberdevices Golos crowd	8.59 %	8.35 %	0.498	0.406
SOVA RuDevices	13.46 %	13.97 %	0.244	0.212
Podlodka Speech	9.61 %	9.27 %	0.383	0.494

WER improves on 4/6 datasets; PER regresses on short read-speech (Golos / SOVA) but improves on long meeting-style speech (Podlodka +0.11, RuLibri +0.06) — the corpus AW was tuned for.

What Stage AW fixed

Stage AS v2 was already an accurate Russian punctuator, but on long same-speaker turns (>15 s) the model would front-load punctuation: periods and commas in the first half, then trail off into punctuation-free run-on text. Empirically 65 % of 20-30 s turns in production meeting logs showed this pattern; users saw walls of text that were correct word-for-word but unreadable.

Stage AW retrains on 611 same-speaker concatenated turns (12-29 s each, 3.88 h total) from Pavel's own meeting corpus, with LLM-restored punctuation targets from gemma3:12b. The recipe continues from Stage AS v2 weights (not from Stage AI) so no short-form quality is sacrificed.

Recipe

• Init: Stage AS v2 merged weights
• Data: 611 rows of same-speaker concat, 12-29 s, gemma3:12b punct-restored targets, WER ≤ 1 % word-preservation verified
• LoRA: r=48, target q/k/v/o + fc1/fc2 (same as AS v2)
• Schedule: 3 epochs, LR 5e-6 cosine, weight decay 0.01
• Compute: ~4 GPU-hours on RTX 4090 mobile

Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

processor = WhisperProcessor.from_pretrained("coriollon/whisper-large-v3-turbo-russian")
model = WhisperForConditionalGeneration.from_pretrained(
    "coriollon/whisper-large-v3-turbo-russian",
    torch_dtype=torch.float16,
).to("cuda")

inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to("cuda", dtype=torch.float16)
predicted_ids = model.generate(input_features, language="ru", task="transcribe")
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

Pre-quantized faster-whisper variants

Folder	Quantization	Size
ct2_int8_float16/	int8 weights + fp16 compute	782 MB

from huggingface_hub import snapshot_download
from faster_whisper import WhisperModel

ct2_path = snapshot_download(
    repo_id="coriollon/whisper-large-v3-turbo-russian",
    allow_patterns="ct2_int8_float16/*",
)
model = WhisperModel(f"{ct2_path}/ct2_int8_float16",
                    device="cuda", compute_type="int8_float16")
segments, _ = model.transcribe("audio.wav", language="ru", beam_size=5)

Reverting to Stage AS v2

The Stage AS v2 revision remains accessible via HF revision pinning:

processor = WhisperProcessor.from_pretrained(
    "coriollon/whisper-large-v3-turbo-russian",
    revision="stage-as-v2",   # git tag pointing to pre-AW commit
)

Limitations

• On short read-speech (Golos / SOVA) PER F1 drops 0.02-0.11 vs AS v2 because AW was tuned for long conversational turns, not short clean utterances. WER is unaffected.
• English-only audio quality inherited from AS v2 baseline — degraded vs vanilla turbo.
• Same over-punctuation risk as AS v2 on very short (< 3 s) inputs.

License

Apache 2.0 (inherited from base Whisper model).

Changelog

• Stage AW (2026-09-07) — long-form density fine-tune. Continue-train from AS v2 on 611 same-speaker concat-turns (12-29 s, gemma3:12b targets, 3.88 h). Long-form density ×2.7-2.9, n_valid 53→140/150, ratio_med 1.14→1.00. Aggregate WER +0.05 pp (in noise); 4/6 datasets improve.
• Metric fix (2026-07-19) — Evaluation harness now counts leading em-dash (Russian dialogue) and standalone ... / … tokens. No change to the released weights, only to how their punctuation quality is reported.
• Stage AS v2 (2026-07-14) — LoRA r=48, 3 epochs, rare-mark ×2 sample weighting. WER 9.83 → 8.93 %. PER 0.39 → 0.41, Cap and Sent essentially preserved. Golos WER drops −2 pp on both splits.
• Stage AS v1 (2026-07-12) — punctuation & capitalization LoRA fine-tune. Cap F1 0.48 → 0.83, Sent F1 0.45 → 0.81, PER F1 0.20 → 0.39. WER 9.43 → 9.83 %.
• Stage AI (previous) — GigaAM-teacher relabel, WER 13.25 → 9.43 %.