task: common_voice_en
task_alias: common_voice_en

# subset only. The real dataset (mozilla-foundation/common_voice_17_0) is to huge
dataset_path: fixie-ai/endpointing-audio
dataset_name: common_voice_17_0-en

training_split: train
test_split: test

output_type: generate_until

doc_to_audio: !function utils.doc_to_audio
doc_to_text: !function utils.doc_to_text
doc_to_target: "{{transcript}}"

generation_kwargs:
  until:
    - "<|endoftext|>"
  temperature: 0.0
  do_sample: false
  max_gen_toks: 64

metric_list:
  - metric: bleu
    aggregation: bleu
    higher_is_better: true

# any metadata you need. Does not affect how the task is processed
metadata:
  version: 1.0