README.md

metadata

language: sv-SE
datasets:
  - common_voice
  - NST Swedish ASR Database
metrics:
  - wer
  - cer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
license: apache-2.0
model-index:
  - name: XLSR Wav2Vec2 Swedish by KBLab
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice sv-SE
          type: common_voice
          args: sv-SE
        metrics:
          - name: Test WER
            type: wer
            value: 14.29861
          - name: Test CER
            type: cer
            value: 4.925294

Wav2Vec2-Large-XLSR-53-Swedish

Fine-tuned facebook/wav2vec2-large-xlsr-53 in Swedish using the NST Swedish Dictation. 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", "sv-SE", split="test[:2%]").

processor = Wav2Vec2Processor.from_pretrained("KBLab/wav2vec2-large-xlsr-53-swedish")
model = Wav2Vec2ForCTC.from_pretrained("KBLab/wav2vec2-large-xlsr-53-swedish")

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 Swedish 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", "sv-SE", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("KBLab/wav2vec2-large-xlsr-53-swedish")
model = Wav2Vec2ForCTC.from_pretrained("KBLab/wav2vec2-large-xlsr-53-swedish")
model.to("cuda")

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"])))
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"]])))

WER: 14.298610% CER: 4.925294%

Training

First the XLSR model was further pre-trained for 50 epochs with a corpus consisting of 1000 hours spoken Swedish from various radio stations. Secondly NST Swedish Dictation was used for fine tuning as well as Common Voice. Lastly only Common Voice dataset was used for final finetuning. The Fairseq scripts were used.