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

Usage

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

python

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

test_dataset = load_dataset("common_voice", "pt", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("gchhablani/wav2vec2-large-xlsr-pt")
model = Wav2Vec2ForCTC.from_pretrained("gchhablani/wav2vec2-large-xlsr-pt")

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["speech"][:2], 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["sentence"][:2])

Evaluation

The model can be evaluated as follows on the Portuguese test data of Common Voice.

python

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

test_dataset = load_dataset("common_voice", "pt", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("gchhablani/wav2vec2-large-xlsr-pt")
model = Wav2Vec2ForCTC.from_pretrained("gchhablani/wav2vec2-large-xlsr-pt")
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\;\"\“\'\�]'
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):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
    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)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def evaluate(batch):
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

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

        pred_ids = torch.argmax(logits, dim=-1)
        batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch

result = test_dataset.map(evaluate, batched=True, batch_size=8)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

Test Result: 17.22 %

Training

The Common Voice train and validation datasets were used for training. The script used for training can be found here. The parameters passed were:

bash

#!/usr/bin/env bash
python run_common_voice.py \
    --model_name_or_path="facebook/wav2vec2-large-xlsr-53" \
    --dataset_config_name="pt" \
    --output_dir=/workspace/output_models/pt/wav2vec2-large-xlsr-pt \
    --cache_dir=/workspace/data \
    --overwrite_output_dir \
    --num_train_epochs="30" \
    --per_device_train_batch_size="32" \
    --per_device_eval_batch_size="32" \
    --evaluation_strategy="steps" \
    --learning_rate="3e-4" \
    --warmup_steps="500" \
    --fp16 \
    --freeze_feature_extractor \
    --save_steps="500" \
    --eval_steps="500" \
    --save_total_limit="1" \
    --logging_steps="500" \
    --group_by_length \
    --feat_proj_dropout="0.0" \
    --layerdrop="0.1" \
    --gradient_checkpointing \
    --do_train --do_eval \

Notebook containing the evaluation can be found here.