README.md

metadata

language:
  - tr
datasets:
  - common_voice
metrics:
  - wer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
license: apache-2.0
model-index:
  - name: XLSR Wav2Vec2 Large Turkish - Gorkem Goknar
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice tr
          type: common_voice
          args: tr
        metrics:
          - name: Test WER
            type: wer
            value: TBD

Wav2Vec2-Large-XLSR-53-Turkish

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Turkish using the Common Voice. 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:

import torch
import torchaudio
import pydub 
from pydub.utils import mediainfo
import array
from pydub import AudioSegment
from pydub.utils import get_array_type
import numpy as np 

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
test_dataset = load_dataset("common_voice", "tr", split="test[:2%]") 
processor = Wav2Vec2Processor.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish")
model = Wav2Vec2ForCTC.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish") 



def audio_resampler(batch,new_sample_rate = 16000):
    
    ##torchaudio and librosa troublesome to use for mp3 in windows
    #speech_array, sampling_rate = torchaudio.load(batch["path"])
    #speech_array, sampling_rate = librosa.load(batch["path"])

    #AudioSegment does the job over ffmpeg(need install)    
    sound = AudioSegment.from_file(file=batch["path"])
    sound = sound.set_frame_rate(new_sample_rate)

    left = sound.split_to_mono()[0]
    bit_depth = left.sample_width * 8
    array_type = get_array_type(bit_depth)

    numeric_array = np.array(array.array(array_type, left._data) )

    #windows hack as torchaudio cannot read mp3
    speech_array = torch.FloatTensor(numeric_array)
    
    batch["speech"] = numeric_array
    batch["sampling_rate"] = new_sample_rate
    batch["target_text"] = batch["sentence"]

    return batch
  
resampler = audio_resampler(16000)

# 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 Turkish test data of Common Voice.

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
test_dataset = load_dataset("common_voice", "tr", split="test") 
wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish") 
model = Wav2Vec2ForCTC.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish") 
model.to("cuda")

#Note: Not ignoring "'"  on this one 
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: TBD %

Training

The Common Voice train and validation datasets were used for training. Additional 5 Turkish movies with subtitles also used