README.md

---
language:
- ru
library_name: transformers
tags:
- asr
- whisper
- russian
datasets:
- mozilla-foundation/common_voice_11_0
metrics:
- wer
---

# Model Details

This is a version of [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) finetuned for better support of Russian language.

Dataset used for finetuning is Common Voice 11.0, Russian part. 

After preprocessing of the original dataset (train + test + validation splits were mixed and split to a new train + test split by 0.95/0.05) the original Whisper v3 has WER 9.2 while the finetuned version shows 6.31 (so far).

## Usage

```
import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor, pipeline

torch_dtype = torch.bfloat16 # set your preferred type here 

device = 'cpu'
if torch.cuda.is_available():
    device = 'cuda'
elif torch.backends.mps.is_available():
    device = 'mps'
    setattr(torch.distributed, "is_initialized", lambda : False) # monkey patching
device = torch.device(device)

whisper = WhisperForConditionalGeneration.from_pretrained(
    "antony66/whisper-large-v3-russian", torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True,
    # add attn_implementation="flash_attention_2" if your GPU supports it
)

processor = WhisperProcessor.from_pretrained("antony66/whisper-large-v3-russian")

asr_pipeline = pipeline(
    "automatic-speech-recognition",
    model=whisper,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=256,
    chunk_length_s=30,
    batch_size=16,
    return_timestamps=True,
    torch_dtype=torch_dtype,
    device=device,
)

# read your wav file into variable wav. For example:
from io import BufferIO
wav = BytesIO()
with open('call.wav', 'rb') as f:
    wav.write(f.read())
wav.seek(0)

# get the transcription
asr = asr_pipeline(wav, generate_kwargs={"language": "russian", "max_new_tokens": 256}, return_timestamps=False)

print(asr['text'])

```

## Work in progress

This model is in WIP state for now. The goal is to finetune it for speech recognition of phone calls as much as possible. If you want to contribute and you know or have any good dataset please let me know. Your help will be much appreciated.