| --- |
| language: |
| - kk |
| license: cc-by-4.0 |
| task_categories: |
| - automatic-speech-recognition |
| tags: |
| - kazakh |
| - medical |
| - aligned |
| - speech |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # Medical ASR Aligned Dataset |
|
|
| Aligned Kazakh medical speech dataset from the «ТЕЛЕДӘРІГЕР» (TeleDoctor) TV program on Qazaqstan National Channel. |
|
|
| ## Dataset Description |
|
|
| Audio-transcript aligned segments of Kazakh-language medical TV broadcasts. Each segment contains the original audio chunk, ASR transcription, human reference transcription, and Character Error Rate (CER). |
|
|
| **Only segments with CER < 25% are included.** |
|
|
| ## Stats |
|
|
| | Split | Segments | Avg CER | Avg Duration | Total Duration | |
| |-------|----------|---------|--------------|----------------| |
| | train | 17,623 | 0.1011 | 15.1s | 74.16h | |
| | dev | 2,017 | 0.1027 | 15.0s | 8.42h | |
| | test | 2,366 | 0.1040 | 15.2s | 9.98h | |
| | **total** | **22,006** | **0.1015** | **15.1s** | **92.55h** | |
|
|
| ## Data Fields |
|
|
| - `audio`: Audio waveform (WAV, mono, 16kHz) |
| - `asr_text`: ASR transcription (Kazakh) |
| - `human_text`: Human reference transcription (Kazakh) |
| - `cer`: Character Error Rate between ASR and human text |
| - `duration`: Segment duration in seconds |
|
|
| ## How the Data Was Aligned |
|
|
| 1. **Source transcripts** were taken from the `transcripts ready` folder, which contained speaker-diarized transcripts from Whisper AI Scribe. |
| 2. **Preprocessing**: metadata headers (Title, Created, Profile, Speakers), timestamp lines, and speaker labels were stripped. The remaining text was concatenated into plain text per episode. |
| 3. **Audio conversion**: all audio files were converted to WAV mono 16kHz. |
| 4. **Alignment**: the EuroSpeech alignment pipeline (`parliament_transcript_aligner`) was used to align ASR output against human transcripts. The pipeline segments audio into 10–20s windows, runs ASR (model: `issai/whisper-tilsync-09oct2025`, language: `kk`), and aligns ASR segments to reference text using CER-based matching. |
| 5. **Filtering**: only segments with CER < 0.25 were retained. |
| 6. **Splitting**: episodes were randomly split 80/10/10 into train/dev/test (seed=42) before alignment. |
|
|
