README.md

---
language: zh
datasets:
- common_voice
metrics:
- cer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Large 53 - Chinese (zh-CN), by Yih-Dar SHIEH
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Common Voice zh-CN
      type: common_voice
      args: zh-CN
    metrics:
       - name: Test CER
         type: cer
         value: 41.99
---

# Wav2Vec2-Large-XLSR-53-Chinese-zh-cn-gpt

Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Chinese (zh-CN) using the [Common Voice](https://huggingface.co/datasets/common_voice), included [Common Voice](https://huggingface.co/datasets/common_voice) Chinese (zh-TW) dataset (converting the label text to simplified Chinese). 
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", "zh-CN", split="test")

processor = Wav2Vec2Processor.from_pretrained("ydshieh/wav2vec2-large-xlsr-53-chinese-zh-cn-gpt")
model = Wav2Vec2ForCTC.from_pretrained("ydshieh/wav2vec2-large-xlsr-53-chinese-zh-cn-gpt")

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 as follows on the zh-CN test data of Common Voice.
Original CER calculation refer to https://huggingface.co/ctl/wav2vec2-large-xlsr-cantonese

```python
!pip install jiwer

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

def chunked_cer(targets, predictions, chunk_size=None):

    _predictions = [char for seq in predictions for char in list(seq)]
    _targets = [char for seq in targets for char in list(seq)]
    
    if chunk_size is None: return jiwer.wer(_targets, _predictions)
    
    start = 0
    end = chunk_size
    H, S, D, I = 0, 0, 0, 0
    
    while start < len(targets):
    
        _predictions = [char for seq in predictions[start:end] for char in list(seq)]
        _targets = [char for seq in targets[start:end] for char in list(seq)]
        chunk_metrics = jiwer.compute_measures(_targets, _predictions)
        H = H + chunk_metrics["hits"]
        S = S + chunk_metrics["substitutions"]
        D = D + chunk_metrics["deletions"]
        I = I + chunk_metrics["insertions"]
        start += chunk_size
        end += chunk_size
        
    return float(S + D + I) / float(H + S + D)
    
test_dataset = load_dataset("common_voice", "zh-CN", split="test")

processor = Wav2Vec2Processor.from_pretrained("ydshieh/wav2vec2-large-xlsr-53-chinese-zh-cn-gpt")
model = Wav2Vec2ForCTC.from_pretrained("ydshieh/wav2vec2-large-xlsr-53-chinese-zh-cn-gpt")
model.to("cuda")

chars_to_ignore_regex = '[\\\\,\\\\?\\\\.\\\\!\\\\-\\\\;\\\\:"\\\\“\\\\%\\\\‘\\\\”\\\\�\\\\．\\\\⋯\\\\！\\\\－\\\\：\\\\–\\\\。\\\\》\\\\,\\\\）\\\\,\\\\？\\\\；\\\\～\\\\~\\\\…\\\\︰\\\\，\\\\（\\\\」\\\\‧\\\\《\\\\﹔\\\\、\\\\—\\\\／\\\\,\\\\「\\\\﹖\\\\·\\\\×\\\\̃\\\\̌\\\\ε\\\\λ\\\\μ\\\\и\\\\т\\\\─\\\\□\\\\〈\\\\〉\\\\『\\\\』\\\\ア\\\\オ\\\\カ\\\\チ\\\\ド\\\\ベ\\\\ャ\\\\ヤ\\\\ン\\\\・\\\\丶\\\\ａ\\\\ｂ\\\\ｆ\\\\ｇ\\\\ｉ\\\\ｎ\\\\ｐ\\\\ｔ' + "\\\\']"

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().replace("’", "'") + " "
    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("CER: {:2f}".format(100 * chunked_cer(predictions=result["pred_strings"], targets=result["sentence"], chunk_size=1000)))
```

**Test Result**: 41.987498 %


## Training

The Common Voice zh-CN `train`, `validation` were used for training, as well as Common Voice zh-TW `train`, `validation` and `test` datasets.

The script used for training can be found [to be uploaded later](...)