README.md

metadata

language: el
datasets:
  - common_voice
  - CSS10 Greek: Single Speaker Speech Dataset
metrics:
  - wer
  - cer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
license: apache-2.0
model-index:
  - name: V XLSR Wav2Vec2 Large 53 - greek
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice el
          type: common_voice
          args: el
        metrics:
          - name: Test WER
            type: wer
            value: 18.996669
          - name: Test CER
            type: cer
            value: 5.781874

Wav2Vec2-Large-XLSR-53-greek

Fine-tuned facebook/wav2vec2-large-xlsr-53 on greek using the Common Voice and CSS10 Greek: Single Speaker Speech 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:

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

test_dataset = load_dataset("common_voice", "el", split="test[:2%]") #TODO: replace {lang_id} in your language code here. Make sure the code is one of the *ISO codes* of [this](https://huggingface.co/languages) site.

processor = Wav2Vec2Processor.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
model = Wav2Vec2ForCTC.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`

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 greek 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", "el", split="test") #TODO: replace {lang_id} in your language code here. Make sure the code is one of the *ISO codes* of [this](https://huggingface.co/languages) site.
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
model = Wav2Vec2ForCTC.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“]' # TODO: adapt this list to include all special characters you removed from the data

normalize_greek_letters = {"ς": "σ"}
# normalize_greek_letters = {"ά": "α", "έ": "ε", "ί": "ι", 'ϊ': "ι", "ύ": "υ", "ς": "σ", "ΐ": "ι", 'ϋ': "υ", "ή": "η", "ώ": "ω", 'ό': "ο"}
remove_chars_greek = {"a": "", "h": "", "n": "", "g": "", "o": "", "v": "", "e": "", "r": "", "t": "", "«": "", "»": "", "m": "", '́': '', "·": "", "’": "", '´': ""}
replacements = {**normalize_greek_letters, **remove_chars_greek}

resampler = {
    48_000: torchaudio.transforms.Resample(48_000, 16_000),
    44100: torchaudio.transforms.Resample(44100, 16_000),
    32000: torchaudio.transforms.Resample(32000, 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()
    for key, value in replacements.items():
        batch["sentence"] = batch["sentence"].replace(key, value)
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler[sampling_rate](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"])))
print("CER: {:2f}".format(100 * wer.compute(predictions=[" ".join(list(entry)) for entry in result["pred_strings"]], references=[" ".join(list(entry)) for entry in result["sentence"]])))

Test Result: 18.996669 %

Training

The Common Voice train dataset was used for training. Also all of CSS10 Greek was used using the normalized transcripts. During text preprocessing letter ς is normalized to σ the reason is that both letters sound the same with ς only used as the ending character of words. So, the change can be mapped up to proper dictation easily. I tried removing all accents from letters as well that improved WER significantly. The model was reaching 17% WER easily without having converged. However, the text preprocessing needed to do after to fix transcrtiptions would be more complicated. A language model should fix things easily though. Another thing that could be tried out would be to change all of ι, η ... etc to a single character since all sound the same. similar for o and ω these should help the acoustic model part significantly since all these characters map to the same sound. But further text normlization would be needed.