comodoro/wav2vec2-xls-r-300m-cs-250

Czech wav2vec2-xls-r-300m-cs-250

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice 8.0 dataset as well as other datasets listed below.

It achieves the following results on the evaluation set:

• —Loss: 0.1271
• —Wer: 0.1475
• —Cer: 0.0329

The eval.py script results using a LM are:

• —WER: 0.07274312090176113
• —CER: 0.021207369275558875

Model description

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Czech using the Common Voice dataset. When using this model, make sure that your speech input is sampled at 16kHz.

The model can be used directly (without a language model) as follows:

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

test_dataset = load_dataset("mozilla-foundation/common_voice_8_0", "cs", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("comodoro/wav2vec2-xls-r-300m-cs-250")
model = Wav2Vec2ForCTC.from_pretrained("comodoro/wav2vec2-xls-r-300m-cs-250")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
	speech_array, sampling_rate = torchaudio.load(batch["path"])
	batch["speech"] = resampler(speech_array).squeeze().numpy()
	return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset[:2]["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
	logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset[:2]["sentence"])

Evaluation

The model can be evaluated using the attached eval.py script:

python eval.py --model_id comodoro/wav2vec2-xls-r-300m-cs-250 --dataset mozilla-foundation/common-voice_8_0 --split test --config cs

Training and evaluation data

The Common Voice 8.0 train and validation datasets were used for training, as well as the following datasets:

• —Šmídl, Luboš and Pražák, Aleš, 2013, OVM – Otázky Václava Moravce, LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University, http://hdl.handle.net/11858/00-097C-0000-000D-EC98-3.

• —Pražák, Aleš and Šmídl, Luboš, 2012, Czech Parliament Meetings, LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University, http://hdl.handle.net/11858/00-097C-0000-0005-CF9C-4.

• —Plátek, Ondřej; Dušek, Ondřej and Jurčíček, Filip, 2016, Vystadial 2016 – Czech data, LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University, http://hdl.handle.net/11234/1-1740.

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 0.0001
• —trainbatchsize: 32
• —evalbatchsize: 8
• —seed: 42
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: linear
• —lrschedulerwarmup_steps: 800
• —num_epochs: 5
• —mixedprecisiontraining: Native AMP

Training results

Framework versions

• —Transformers 4.16.2
• —Pytorch 1.10.1+cu102
• —Datasets 1.18.3
• —Tokenizers 0.11.0