Mining Parliamentary Gold - Building Hebrew ASR
summary
Shows a reproducible Python pipeline that remuxes with light ffmpeg, uses stable_ts + faster-whisper + CTranslate2 to force-align noisy Knesset audio and extract sentence timestamps despite heckling, diglossia and wobbling timestamps. Describes an adaptive alignment algorithm that recovers segments from long sessions, salvages roughly 60% usable and about 40% high quality of thousands of hours, yields a license-clean Hugging Face dataset and produces a measured 10% WER reduction when fine-tuning Whisper variants.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:07Hebrew Hi project
- 2:34Knesset plenary data
- 5:30Data technical aspects
- 10:54Training with data
- 11:59Call to action
related talks
01h 34mReinventing Speech-to-Text Transcriptions with Go and Whisper
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
20mVoice fingerprinting & recognition
31m“Sorry, Could you repeat that again?” - Speech Recognition with Python
16mThe End of Awkward AI Transcriptions
25mBeyond Transcription: Building Voice AI That Understands Conversations
20mMeta's AI Translations Inference
59mTechnical Workshop: Sarcastically Speaking: Unlocking Multi-modal Sentiment Analysis With NLP and Facial Expressions