Perch publishes three open releases for checking drug names in dictation

Описание и применение

Perch выпустил три открытых ресурса для проверки названий лекарств в диктовке: инструмент оценки фонем, таблицу произношений и наборы скрытых диктовок, позволяющие воспроизводить и улучшать распознавание речи в ветеринарной практике.

Задача

Existing dictation systems misrecognize veterinary drug names, making accurate transcription difficult.

Результат

Perch’s releases provide reproducible, open resources for evaluating veterinary drug name recognition, facilitating community testing and improvement of ASR systems.

Рабочий процесс

• 1. Run phoneme lane on dictation audio to score drug names against phoneme model.
• 2. Use vet drug pronunciations table to map names to IPA for American, British, and South Asian English.
• 3. Test ASR systems against blind dictation corpora (70 rows each, 45 dictations + 25 traps).
• 4. Compare ASR outputs (Whisper large v3, MedASR) to ground truth and identify misrecognitions.
• 5. Report misses via issue on corresponding GitHub repo.

Доказательства

The phoneme lane scores the audio, not the transcript.

Источник

Vet drug pronunciations lists 265 drug and brand names used in veterinary practice, each with its pronunciation in IPA.

Источник

Blind dictation corpora are six sets of 70 rows, each a dictation or a trap, each written by an author who never saw Perch's matching code.

Источник