jonatasgrosman's picture
first commit
64967ea
metadata
language: en
datasets:
  - common_voice
metrics:
  - wer
  - cer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
license: apache-2.0
model-index:
  - name: XLSR Wav2Vec2 English by Jonatas Grosman
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice en
          type: common_voice
          args: en
        metrics:
          - name: Test WER
            type: wer
            value: 39.59
          - name: Test CER
            type: cer
            value: 18.18

Wav2Vec2-Large-XLSR-53-English

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

The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint

Usage

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

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

LANG_ID = "en"
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-english"
SAMPLES = 10

test_dataset = load_dataset("common_voice", LANG_ID, split=f"test[:{SAMPLES}]")

processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
    batch["speech"] = speech_array
    batch["sentence"] = batch["sentence"].upper()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["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)
predicted_sentences = processor.batch_decode(predicted_ids)

for i, predicted_sentence in enumerate(predicted_sentences):
    print("-" * 100)
    print("Reference:", test_dataset[i]["sentence"])
    print("Prediction:", predicted_sentence)
Reference Prediction
"SHE'LL BE ALL RIGHT." SHE'LD BE ALL RIGHT
SIX SIX
"ALL'S WELL THAT ENDS WELL." ALL IS WELL THAT ENDS WELL
DO YOU MEAN IT? DO YOU MEAN IT
THE NEW PATCH IS LESS INVASIVE THAN THE OLD ONE, BUT STILL CAUSES REGRESSIONS. THE NEW PATCH IS LESS INVASIVE THAN THE OLD ONE BUT STILL CAUSES REGRESSION
HOW IS MOZILLA GOING TO HANDLE AMBIGUITIES LIKE QUEUE AND CUE? HOWIS MOCILE ARE GOING TO HANDLE AMBIGUITIES LIKE KU AND KU
"I GUESS YOU MUST THINK I'M KINDA BATTY." RISSHON WAS INCAN IN THE BAK TE
NO ONE NEAR THE REMOTE MACHINE YOU COULD RING? NO ONE NEAR THE REMOTE MACHINE YOU COULD RING
SAUCE FOR THE GOOSE IS SAUCE FOR THE GANDER. SAUCE FOR THE GUISE IS SAUCED FOR THE GONDER
GROVES STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD. GRAFS STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD

Evaluation

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

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

LANG_ID = "en"
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-english"
DEVICE = "cuda"

CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
                   "、", "﹂", "﹁", "‧", "~", "﹏", ",", "{", "}", "(", ")", "[", "]", "【", "】", "‥", "〽",
                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", ":", "!", "?", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]

test_dataset = load_dataset("common_voice", LANG_ID, split="test")

# uncomment the following lines to eval using other datasets
# test_dataset = load_dataset("librispeech_asr", "clean", split="test")
# test_dataset = load_dataset("librispeech_asr", "other", split="test")
# test_dataset = load_dataset("timit_asr", split="test")

wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py

chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"

processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
model.to(DEVICE)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        speech_array, sampling_rate = librosa.load(batch["file"] if "file" in batch else batch["path"], sr=16_000)
    batch["speech"] = speech_array
    batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["text"] if "text" in batch else batch["sentence"]).upper()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the audio 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(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).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)

predictions = [x.upper() for x in result["pred_strings"]]
references = [x.upper() for x in result["sentence"]]

print(f"WER: {wer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")
print(f"CER: {cer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")

Test Result:

In the table below I report the Word Error Rate (WER) and the Character Error Rate (CER) of the model. I ran the evaluation script described above on other models as well (on 2021-05-20). Note that the table below may show different results from those already reported, this may have been caused due to some specificity of the other evaluation scripts used.


Common Voice

Model WER CER
jonatasgrosman/wav2vec2-large-xlsr-53-english 19.18% 8.25%
jonatasgrosman/wav2vec2-large-english 21.16% 9.53%
facebook/wav2vec2-large-960h-lv60-self 22.03% 10.39%
facebook/wav2vec2-large-960h-lv60 23.97% 11.14%
facebook/wav2vec2-large-960h 32.79% 16.03%
boris/xlsr-en-punctuation 34.81% 15.51%
facebook/wav2vec2-base-960h 39.86% 19.89%
facebook/wav2vec2-base-100h 51.06% 25.06%
elgeish/wav2vec2-large-lv60-timit-asr 59.96% 34.28%
facebook/wav2vec2-base-10k-voxpopuli-ft-en 66.41% 36.76%
elgeish/wav2vec2-base-timit-asr 68.78% 36.81%

LibriSpeech (clean)

Model WER CER
facebook/wav2vec2-large-960h-lv60-self 1.86% 0.54%
facebook/wav2vec2-large-960h-lv60 2.15% 0.61%
facebook/wav2vec2-large-960h 2.82% 0.84%
facebook/wav2vec2-base-960h 3.44% 1.06%
facebook/wav2vec2-base-100h 6.26% 2.00%
jonatasgrosman/wav2vec2-large-xlsr-53-english 6.97% 2.02%
jonatasgrosman/wav2vec2-large-english 8.00% 2.55%
elgeish/wav2vec2-large-lv60-timit-asr 15.53% 4.93%
boris/xlsr-en-punctuation 19.28% 6.45%
elgeish/wav2vec2-base-timit-asr 29.19% 8.38%
facebook/wav2vec2-base-10k-voxpopuli-ft-en 31.82% 12.41%

LibriSpeech (other)

Model WER CER
facebook/wav2vec2-large-960h-lv60-self 3.89% 1.40%
facebook/wav2vec2-large-960h-lv60 4.45% 1.56%
facebook/wav2vec2-large-960h 6.49% 2.52%
facebook/wav2vec2-base-960h 8.90% 3.55%
jonatasgrosman/wav2vec2-large-xlsr-53-english 11.75% 4.23%
jonatasgrosman/wav2vec2-large-english 13.62% 5.24%
facebook/wav2vec2-base-100h 13.97% 5.51%
boris/xlsr-en-punctuation 26.40% 10.11%
elgeish/wav2vec2-large-lv60-timit-asr 28.39% 12.08%
elgeish/wav2vec2-base-timit-asr 42.04% 15.57%
facebook/wav2vec2-base-10k-voxpopuli-ft-en 45.19% 20.32%

TIMIT

Model WER CER
facebook/wav2vec2-large-960h-lv60-self 5.17% 1.33%
facebook/wav2vec2-large-960h-lv60 6.24% 1.54%
facebook/wav2vec2-large-960h 9.63% 2.19%
facebook/wav2vec2-base-960h 11.48% 2.76%
jonatasgrosman/wav2vec2-large-xlsr-53-english 11.93% 3.50%
elgeish/wav2vec2-large-lv60-timit-asr 13.83% 4.36%
jonatasgrosman/wav2vec2-large-english 13.91% 4.01%
facebook/wav2vec2-base-100h 16.75% 4.79%
elgeish/wav2vec2-base-timit-asr 25.40% 8.16%
boris/xlsr-en-punctuation 25.93% 9.99%
facebook/wav2vec2-base-10k-voxpopuli-ft-en 51.08% 19.84%