wav2vec2-large-xlsr-53-icelandic-ep30-967h

The fine-tuning process was performed during July (2023) in the servers of the Language and Voice Laboratory (https://lvl.ru.is/) at Reykjavík University (Iceland) by Carlos Daniel Hernández Mena.

Evaluation

import torch
from transformers import Wav2Vec2Processor
from transformers import Wav2Vec2ForCTC

#Load the processor and model.
MODEL_NAME="language-and-voice-lab/wav2vec2-large-xlsr-53-icelandic-ep30-967h"
processor = Wav2Vec2Processor.from_pretrained(MODEL_NAME)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_NAME)

#Load the dataset
from datasets import load_dataset, load_metric, Audio
ds=load_dataset("language-and-voice-lab/samromur_children", split="test")

#Downsample to 16kHz
ds = ds.cast_column("audio", Audio(sampling_rate=16_000))

#Process the dataset
def prepare_dataset(batch):
    audio = batch["audio"]
    #Batched output is "un-batched" to ensure mapping is correct
    batch["input_values"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
    with processor.as_target_processor():
        batch["labels"] = processor(batch["normalized_text"]).input_ids
    return batch
ds = ds.map(prepare_dataset, remove_columns=ds.column_names,num_proc=1)

#Define the evaluation metric
import numpy as np
wer_metric = load_metric("wer")
def compute_metrics(pred):
    pred_logits = pred.predictions
    pred_ids = np.argmax(pred_logits, axis=-1)
    pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id
    pred_str = processor.batch_decode(pred_ids)
    #We do not want to group tokens when computing the metrics
    label_str = processor.batch_decode(pred.label_ids, group_tokens=False)
    wer = wer_metric.compute(predictions=pred_str, references=label_str)
    return {"wer": wer}

#Do the evaluation (with batch_size=1)
model = model.to(torch.device("cuda"))
def map_to_result(batch):
    with torch.no_grad():
        input_values = torch.tensor(batch["input_values"], device="cuda").unsqueeze(0)
        logits = model(input_values).logits
    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_str"] = processor.batch_decode(pred_ids)[0]
    batch["sentence"] = processor.decode(batch["labels"], group_tokens=False)
    return batch
results = ds.map(map_to_result,remove_columns=ds.column_names)

#Compute the overall WER now.
print("Test WER: {:.3f}".format(wer_metric.compute(predictions=results["pred_str"], references=results["sentence"])))

Test Result: 0.076

BibTeX entry and citation info

When publishing results based on these models please refer to:

@inproceedings{mena2024samromur,
  title={Samr{\'o}mur Millj{\'o}n: An ASR Corpus of One Million Verified Read Prompts in Icelandic},
  author={Mena, Carlos Daniel Hernandez and Gunnarsson, {\TH}orsteinn Da{\dh}i and Gu{\dh}nason, J{\'o}n},
  booktitle={Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  pages={14305--14312},
  year={2024}
}

Acknowledgements

Thanks to Jón Guðnason, head of the Language and Voice Lab for providing computational power to make this model possible.

We also want to thank to the "Language Technology Programme for Icelandic 2019-2023" which is managed and coordinated by Almannarómur, and it is funded by the Icelandic Ministry of Education, Science and Culture. This model is an unexpected result of all the resources gathered by the Programme.

Special thanks to Björn Ingi Stefánsson for setting up the configuration of the server where this model was trained.