test_hf_dataset

This dataset was created to document how to create an audio dataset and upload it to HuggingFace see GitHub repo.

Next step: add more documentation. e.g.:

• contents of the dataset
• context for how the dataset should be used, e.g.: datasets package
• existing dataset cards, such as the ELI5 dataset card, show common conventions

Example usage of dataset

Example of transcription.

First install extra dependencies, typically within virtual environment.

python3 -m pip install datasets torch transformers

Then save and run this Python script. It runs transcription using the Moonshine model by Useful Sensors link.

"""Adapted from https://github.com/usefulsensors/moonshine#huggingface-transformers"""
from datasets import load_dataset
from transformers import AutoProcessor, MoonshineForConditionalGeneration

dataset = load_dataset("guynich/test_hf_dataset", split="test")
model = MoonshineForConditionalGeneration.from_pretrained(
    "UsefulSensors/moonshine-tiny"
)
processor = AutoProcessor.from_pretrained("UsefulSensors/moonshine-tiny")

for index in range(len(dataset)):
    audio_array = dataset[index]["audio"]["array"]
    sampling_rate = dataset[index]["audio"]["sampling_rate"]

    inputs = processor(audio_array, return_tensors="pt", sampling_rate=sampling_rate)

    generated_ids = model.generate(**inputs)

    transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
    print(transcription)

Example output.

$ python3 main.py
The birch canoe slid on the smooth planks, glue the sheets to a dark blue background.
$