---
license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: record_id
    dtype: string
  - name: speciality
    dtype: string
  - name: name
    dtype: string
  - name: description
    dtype: string
  - name: chunk_index
    dtype: int64
  - name: transcription
    dtype: string
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: duration_s
    dtype: float64
  - name: voice
    dtype: string
  - name: speed
    dtype: float64
  - name: lang_code
    dtype: string
  - name: keywords
    list: string
  splits:
  - name: train
    num_bytes: 25738669262.636265
    num_examples: 39569
  - name: validation
    num_bytes: 3217252348.3787556
    num_examples: 4946
  - name: test
    num_bytes: 3217902823.984978
    num_examples: 4947
  download_size: 30140033145
  dataset_size: 32173824435.0
language:
- en
tags:
- medical
- kokoro
- asr
- audio
size_categories:
- 10K<n<100K
---
# MTSamples‑Kokoro‑ASR (Synthetic Medical Speech)

**Summary:** 279 hours of **synthetic** English medical speech (49,462 clips) created from publicly available transcripts on **MTSamples.com** using multiple US/UK voices from **Kokoro‑82M**. Intended for training and evaluating **medical ASR**.

## Dataset
- **Rows:** 49,462  
- **Total audio:** ~279 hours (mono)  
- **Source text:** Sample medical reports from MTSamples.com (names/dates typically altered or removed)  
- **Audio generation:** hexgrad/**Kokoro‑82M** (various American/British voices)  
- **Primary use:** Supervised training/fine‑tuning of medical ASR (e.g., Whisper, Conformer, wav2vec 2.0)

## Features (columns)
- `record_id` — unique ID for the source record/sample  
- `speciality` — medical specialty/category (e.g., Cardiology, Orthopedics)  
- `name` — short title of the report/sample  
- `description` — brief summary (if available)  
- `chunk_index` — 0‑based index of this chunk within the sample  
- `transcription` — exact text used to synthesize this row’s audio  
- `audio` — `datasets.Audio` object (16000 Hz)  
- `duration_s` — duration of the audio in seconds  
- `voice` — Kokoro voice tag used for synthesis  
- `speed` — relative speaking rate used for synthesis (1.0 = default)  
- `lang_code` — language/locale code for the selected voice (e.g., en‑US, en‑GB)  
- `keywords` — optional keywords extracted from `chunk_text`

## Quick start (Hugging Face Datasets)
```python
from datasets import load_dataset, Audio

ds = load_dataset("YOUR_ORG/MTSamples-Kokoro-ASR", split="train")

ex = ds[0]
print(ex["record_id"], ex["speciality"], ex["voice"], ex["duration_s"])
```

### Tips
- When creating train/dev/test, **group by `record_id`** to avoid chunk leakage across splits.
- Use `chunk_text` as the ASR target; consider balancing by `voice` and `speed` for robustness.

## Licensing & Use Notes
- **Text:** Derived from MTSamples.com; review their site notices/terms for redistribution or commercial use.  
- **Audio:** Synthetic speech generated from Kokoro‑82M; follow its license.  
- This card is informational and **not legal advice**.

## Acknowledgements

- hexgrad. **Kokoro‑82M: English TTS** (model). Hugging Face. https://huggingface.co/hexgrad/Kokoro-82M