README.md

metadata

language:
  - he
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
  - text-to-speech
pretty_name: Knesset VOX IPA
size_categories:
  - 10K<n<100K
tags:
  - hebrew
  - ipa
  - phonemes
  - knesset
  - parliament
  - speech

Knesset VOX IPA

Hebrew speech dataset derived from Knesset (Israeli Parliament) plenary sessions, enriched with IPA (International Phonetic Alphabet) phoneme transcriptions. Inspired by the methodology of arxiv:2603.01270.

Dataset Description

Long-form Knesset recordings were split into chunks of up to 15 seconds. Each chunk was transcribed to Hebrew text and then processed for IPA phoneme extraction from audio.

Each sample pairs a WAV audio chunk with:

• The original Hebrew text transcript
• An IPA phoneme transcription extracted from audio

Dataset Stats

Feature	Value
Language	Hebrew (he)
Total samples	49,433
Audio format	WAV
Max chunk duration	15 seconds
STT model	ivrit-ai/whisper-large-v3-turbo-ct2
IPA model	thewh1teagle/whisper-heb-ipa-large-v3-turbo-ct2

Dataset Structure

Files

• ipa_voxknesset.7z — archive containing all WAV audio chunks
• metadata_ipa.csv — CSV with columns: filename, text, phonemes
• checkpoint_ipa.json — processing checkpoint with per-file metadata

Data Fields

Column	Type	Description
filename	string	WAV chunk file name (e.g. 433_2120_10152_10474.wav)
text	string	Hebrew transcript of the utterance
phonemes	string	IPA phoneme sequence extracted from audio

Example

filename: 433_2120_10152_10474.wav
text: אדוני ראש הממשלה, חברי השרים...
phonemes: ʔadˈonˈi ʁˈoʃ hamemʃalˈa, χaveʁˈaj hasaʁˈim...

Methodology

1. Source audio from Knesset VOX parliamentary recordings
2. Split recordings into chunks of up to 15 seconds
3. Transcribe each chunk to Hebrew text using ivrit-ai/whisper-large-v3-turbo-ct2
4. Extract IPA phonemes from audio using thewh1teagle/whisper-heb-ipa-large-v3-turbo-ct2

Uses

• Hebrew TTS / ASR model training
• Grapheme-to-phoneme (G2P) model training
• Linguistic research on Modern Hebrew phonology