Specialized variant of coriollon/whisper-large-v3-turbo-russian optimized for Russian speech with embedded English tech terms ("Открой Python и сделай git push").

When to use this vs the standard model

Use case	Recommended model
Pure Russian dictation	coriollon/whisper-large-v3-turbo-russian (9.43% AGG WER)
RU+EN code-switching (IT dictation, tech podcasts)	this model

This variant trades ~2 pp on pure Russian WER for a massive code-switching gain.

Code-switching benchmark (540 RU sentences with EN tech terms)

Metric	Standard RU model	This model	Δ
WER	24.64%	9.08%	−15.56 pp
Term accuracy¹	12.59%	74.26%	+61.67 pp

¹ Term accuracy = fraction of utterances where the English tech term (e.g. "Python", "GitHub", "Docker") appears verbatim in latin script in the transcript.

Pure-Russian regression

Dataset	Standard RU	This model	Δ
Common Voice 21 RU	5.33%	6.87%	+1.54
RuLibriSpeech	8.50%	8.57%	+0.07
Sberdevices Golos farfield	11.05%	10.38%	−0.67
Sberdevices Golos crowd	10.15%	9.81%	−0.34
SOVA RuDevices	13.61%	14.87%	+1.26
Podlodka Speech	10.85%	(noisy n=20)	—
AGG (N=200)	~9.4%	~11.8%	+2.3

Usage

python

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

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

inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")
feats = inputs.input_features.to("cuda", dtype=torch.float16)
ids = model.generate(feats, language="ru", task="transcribe", num_beams=5)
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

Pre-quantized faster-whisper variant

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

python

from huggingface_hub import snapshot_download
from faster_whisper import WhisperModel

ct2_path = snapshot_download(
    repo_id="coriollon/whisper-large-v3-turbo-russian-codeswitch",
    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)

How it was built

LoRA r=32 (q+k+v+o+fc1+fc2, 27.85M params) trained on:

2,160 RU sentences with EN tech terms (Silero TTS-bootstrapped, 50 templates × 53 terms × 5 voices)
5,000 pure-RU anti-forgetting samples
2,000 EN anti-forgetting samples

Reference labels in mixed script ("Открой Python") teach the model to natively output latin tokens for English terms instead of cyrillic transliteration ("питон").

Limitations

Pure-RU WER is ~2 pp worse than the standard model. Don't use this model if you don't need code-switching.
Training data was TTS-generated (Silero v4 RU) — real human pronunciation may differ slightly
Currently covers 53 common tech/brand terms; rare terms may still appear in cyrillic

License

Apache 2.0 (inherited from base Whisper model).