okuparinen/SKN_300m_detailed

Detailed automatic dialectal transcription of Finnish

This is a fine-tuned model for automatic dialectal transcription of Finnish dialect recordings. The model is based on a model trained on colloquial Finnish: GetmanY1/wav2vec2-large-fi-lp-cont-pt. The model has been finetuned on old Finnish dialect recordings and their corresponding transcriptions in the Uralic Phonetic Alphabet. This model outputs detailed transcription. The audio recordings are sampled at 16kHz.

Uses

You can use this model for automatic dialectal transcription of Finnish dialects. Note that this model does not produce standard Finnish text.

How to Get Started with the Model

Use the code below to get transcriptions for local data in csv format. If using data from the HuggingFace Hub, see load_dataset.

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC, Wav2Vec2CTCTokenizer
from datasets import Dataset, Audio
import torch
import pandas as pd

ds = pd.read_csv('CSV_DATA.csv')
ds = ds.dropna(how='any', axis=0)

test = Dataset.from_pandas(skn_test)
test = test.cast_column("AUDIO_PATH_COLUMN", Audio(sampling_rate=16000))

tokenizer = Wav2Vec2CTCTokenizer.from_pretrained("okuparinen/SKN_300m_detailed", unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")
model = Wav2Vec2ForCTC.from_pretrained("okuparinen/SKN_300m_detailed").to("cuda")
processor = Wav2Vec2Processor.from_pretrained("okuparinen/SKN_300m_detailed", tokenizer=tokenizer)

def prepare_dataset(batch):
    audio = batch["AUDIO_PATH"]
    batch["input_values"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
    batch["input_length"] = len(batch["input_values"])

    return batch

test_ready = test.map(prepare_dataset, remove_columns=test.column_names)

length = len(test)
predictions = []

for i in range(0, length, 1):
    input_dict = processor(test_ready[i]["input_values"], return_tensors="pt", padding=True)
    logits = model(input_dict.input_values.to("cuda")).logits
    
    pred_ids = torch.argmax(logits, dim=-1)[0]
    
    prediction = processor.decode(pred_ids)
    predictions.append(prediction)

with open("OUTFILE.txt", "w") as f_pred:
    for line in predictions:
        f_pred.write(line + '\n')

Training Data

The training data is an utterance-level version of the Samples of Spoken Finnish corpus. The utterance-level version is available at okuparinen/skn.

Evaluation results

TBA

Citation

BibTeX:

[More Information Needed]