Edit model card


The Wav2Vec2 model is based on facebook/wav2vec2-large-xlsr-53, fine-tuned in Russian using Sberdevices Golos with audio augmentations like as pitch shift, acceleration/deceleration of sound, reverberation etc.

When using this model, make sure that your speech input is sampled at 16kHz.


To transcribe audio files the model can be used as a standalone acoustic model as follows:

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torch
# load model and tokenizer
processor = Wav2Vec2Processor.from_pretrained("bond005/wav2vec2-large-ru-golos")
model = Wav2Vec2ForCTC.from_pretrained("bond005/wav2vec2-large-ru-golos")
# load the test part of Golos dataset and read first soundfile
ds = load_dataset("bond005/sberdevices_golos_10h_crowd", split="test")
# tokenize
processed = processor(ds[0]["audio"]["array"], return_tensors="pt", padding="longest")  # Batch size 1
# retrieve logits
logits = model(processed.input_values, attention_mask=processed.attention_mask).logits
# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]


This code snippet shows how to evaluate bond005/wav2vec2-large-ru-golos on Golos dataset's "crowd" and "farfield" test data.

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torch
from jiwer import wer, cer  # we need word error rate (WER) and character error rate (CER)

# load the test part of Golos Crowd and remove samples with empty "true" transcriptions
golos_crowd_test = load_dataset("bond005/sberdevices_golos_10h_crowd", split="test")
golos_crowd_test = golos_crowd_test.filter(
    lambda it1: (it1["transcription"] is not None) and (len(it1["transcription"].strip()) > 0)

# load the test part of Golos Farfield and remove sampels with empty "true" transcriptions
golos_farfield_test = load_dataset("bond005/sberdevices_golos_100h_farfield", split="test")
golos_farfield_test = golos_farfield_test.filter(
    lambda it2: (it2["transcription"] is not None) and (len(it2["transcription"].strip()) > 0)

# load model and tokenizer
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h").to("cuda")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")

# recognize one sound
def map_to_pred(batch):
    # tokenize and vectorize
    processed = processor(
        batch["audio"]["array"], sampling_rate=batch["audio"]["sampling_rate"],
        return_tensors="pt", padding="longest"
    input_values = processed.input_values.to("cuda")
    attention_mask = processed.attention_mask.to("cuda")

    # recognize
    with torch.no_grad():
        logits = model(input_values, attention_mask=attention_mask).logits
    predicted_ids = torch.argmax(logits, dim=-1)

    # decode
    transcription = processor.batch_decode(predicted_ids)
    batch["text"] = transcription[0]
    return batch

# calculate WER and CER on the crowd domain
crowd_result = golos_crowd_test.map(map_to_pred, remove_columns=["audio"])
crowd_wer = wer(crowd_result["transcription"], crowd_result["text"])
crowd_cer = cer(crowd_result["transcription"], crowd_result["text"])
print("Word error rate on the Crowd domain:", crowd_wer)
print("Character error rate on the Crowd domain:", crowd_cer)

# calculate WER and CER on the farfield domain
farfield_result = golos_farfield_test.map(map_to_pred, remove_columns=["audio"])
farfield_wer = wer(farfield_result["transcription"], farfield_result["text"])
farfield_cer = cer(farfield_result["transcription"], farfield_result["text"])
print("Word error rate on the Farfield domain:", farfield_wer)
print("Character error rate on the Farfield domain:", farfield_cer)

Result (WER, %):

"crowd" "farfield"
10.144 20.353

Result (CER, %):

"crowd" "farfield"
2.168 6.030

You can see the evaluation script on other datasets, including Russian Librispeech and SOVA RuDevices, on my Kaggle web-page https://www.kaggle.com/code/bond005/wav2vec2-ru-eval


If you want to cite this model you can use this:

  title={XLSR Wav2Vec2 Russian by Ivan Bondarenko},
  author={Bondarenko, Ivan},
  publisher={Hugging Face},
  journal={Hugging Face Hub},
Downloads last month
Hosted inference API
or or
This model can be loaded on the Inference API on-demand.

Datasets used to train bond005/wav2vec2-large-ru-golos

Space using bond005/wav2vec2-large-ru-golos 1

Evaluation results