README.md

metadata

language: zh-CN
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 Chinese (zh-CN) by Jonatas Grosman
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice zh-CN
          type: common_voice
          args: zh-CN
        metrics:
          - name: Test WER
            type: wer
            value: 85.1
          - name: Test CER
            type: cer
            value: 20.16

Wav2Vec2-Large-XLSR-53-Chinese-zh-CN

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Chinese using the Common Voice and CSS10. 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 = "zh-CN"
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-chinese-zh-cn"
SAMPLES = 5

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
宋朝末年年间定居粉岭围。	宋朝末年年间定区粉令为
渐渐行动不便	建震行动变
二十一年去世。	二十一年去世
他们自称恰哈拉。	他们自称恰哈拉
局部干涩的例子包括有口干、眼睛干燥、及阴道干燥。	局物干色的例子包括有口干眼睛干燥以及音到干燥

Evaluation

The model can be evaluated as follows on the Chinese (zh-CN) 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 = "zh-CN"
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-chinese-zh-cn"
DEVICE = "cuda"

CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞", 
                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
                   "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。"]

test_dataset = load_dataset("common_voice", LANG_ID, 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["path"], sr=16_000)
    batch["speech"] = speech_array
    batch["sentence"] = re.sub(chars_to_ignore_regex, "", 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):
\tinputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

\twith torch.no_grad():
\t\tlogits = model(inputs.input_values.to(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).logits

\tpred_ids = torch.argmax(logits, dim=-1)
\tbatch["pred_strings"] = processor.batch_decode(pred_ids)
\treturn 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:

My model may report better scores than others because of some specificity of my evaluation script, so I ran the same evaluation script on other models (on 2021-04-21) to make a fairer comparison.

Model	WER	CER
jonatasgrosman/wav2vec2-large-xlsr-53-chinese-zh-cn	85.10%	20.16%
ydshieh/wav2vec2-large-xlsr-53-chinese-zh-cn-gpt	84.45%	21.32%