A filtered (<=30s duration) slice (512 samples) of the Earnings22 dataset.
def add_duration(sample):
y, sr = sample['audio']["array"], sample['audio']["sampling_rate"]
sample['duration_ms']=librosa.get_duration(y=y, sr=sr) * 1000
return sample
tedlium = load_dataset("esb/datasets", "earnings22", split='validation', trust_remote_code=True)
compute duration to filter
tedlium = tedlium.map(add_duration)
tedlium = tedlium.select(range(512))
Whisper max supported duration
