---
language: en
tags:
  - automatic-speech-recognition
  - medical
  - preprocessing
  - whisper
  - wav2vec2
  - huggingface
license: apache-2.0
---

# prepared_medical_speech

## Dataset Summary

`prepared_medical_speech` holds the Hugging Face `Hani89/medical_asr_recording_dataset` split into train/validation/test subsets with normalized transcripts and mono 16 kHz waveforms that both Whisper and Wav2Vec2 pipelines share. Each sample pairs a floating-point waveform array with a lowercased, whitespace-normalized medical sentence covering symptoms, conditions, and complaints.

## Supported Tasks and Leaderboards

- Task: Automatic Speech Recognition (ASR)
- Models: Whisper (encoder-decoder) & Wav2Vec2 (CTC)
- Intended use: fine-tuning, evaluation, and contextual prompting benchmarks targeting medical vocabulary

## Languages

- English (medical, symptom-focused)

## Dataset Structure

Each split contains examples with two fields:

- `audio`: dictionary with `array` (list of float samples), `sampling_rate` (int, 16 000), and `path` (optional)
- `sentence`: normalized transcript (lowercased, collapsed whitespace)

Splits:

- `train`: used for fine-tuning (after carving 10 % for validation)
- `validation`: sampled from the original training split (seeded split for reproducibility)
- `test`: preserved Hugging Face test split (unchanged content)

## Data Collection

Derived from the publicly available [Medical Speech, Transcription, and Intent dataset](https://huggingface.co/datasets/Hani89/medical_asr_recording_dataset). Audio and transcripts are untouched except for resampling/reshaping and standardized transcript casing.

## Preprocessing

1. Cast every sample to mono and assert 16 kHz sampling rate to satisfy Whisper/Wav2Vec2 requirements.
2. Normalize `sentence` by trimming, lowercasing, and squashing extra whitespace.
3. Split the original training split into `train` vs `validation` with a 90/10 split (seed 42) while preserving Hugging Face `test`.
4. Save the resulting `DatasetDict` to `prepared_medical_speech` for reuse.

## Provenance

Source: `Hani89/medical_asr_recording_dataset` on Hugging Face (dataset includes about 8.5 hours, 6,661 utterances).

## Licensing

Original dataset: import license from Hugging Face (check `Hani89/medical_asr_recording_dataset`). This prepared split inherits the same terms; confirm any restrictions before redistribution.

## Citation

If you use this prepared data, cite the original dataset authors via the Hugging Face page and mention that it was preprocessed for Whisper/Wav2Vec2 experiments.