whisper-small-nl

This model is a fine-tuned version of qmeeus/whisper-small-nl on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3034
• Wer: 14.5354

Model description

More information needed

Intended uses & limitations

Transcribe files in Dutch:

import soundfile as sf
from transformers import pipeline

whisper_asr = pipeline("automatic-speech-recognition", model="qmeeus/whisper-small-nl", device=0)
whisper_asr.model.config.forced_decoder_ids = whisper_asr.tokenizer.get_decoder_prompt_ids(
    task="transcribe", language="nl"
)

waveform, sr = sf.read(filename)

def iter_chunks(waveform, sampling_rate=16_000, chunk_length=30.):
    assert sampling_rate == 16_000
    n_frames = math.floor(sampling_rate * chunk_length)
    for start in range(0, len(waveform), n_frames):
        end = min(len(waveform), start + n_frames)
        yield waveform[start:end]

for sentence in whisper_asr(iter_chunks(waveform, sr), max_new_tokens=448):
    print(sentence["text"])

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2045	2.49	1000	0.3194	16.1628
0.0652	4.97	2000	0.3425	16.3672
0.0167	7.46	3000	0.3915	15.8187
0.0064	9.95	4000	0.4190	15.7298
0.1966	2.02	5000	0.3298	15.0881
0.1912	4.04	6000	0.3266	14.8764
0.1008	7.02	7000	0.3261	14.8086
0.0899	9.04	8000	0.3196	14.6487
0.1126	12.02	9000	0.3283	14.5894
0.1071	14.04	10000	0.3034	14.5354

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2