Show HN: Transcribe-Critic – Merge transcript sources for stronger transcript (github.com)

Transcribe-Critic is an innovative automated pipeline designed to enhance the accuracy of speech transcripts from video content. By utilizing multiple Whisper models, along with YouTube captions and optional external transcripts, it merges these diverse sources into a single, highly precise "critical text." This process applies principles from textual criticism, where an LLM adjudicates discrepancies between the transcripts without bias towards any single source, effectively treating them as independent witnesses to the same speech. The method showcases significant advancements over existing models, like WhisperX, by employing a multi-model ensemble approach that balances accuracy and speed through meticulous chunk alignment and blind adjudication of differences. The significance of Transcribe-Critic for the AI/ML community lies in its advanced integration of LLM adjudication into the transcription process, drastically improving the quality of generated transcripts. With the capability to process 2-3+ sources and retain structured data like speaker labels and timestamps, it provides a detailed and organized output that is not only useful for accurate transcription but also for generating summaries and analyzing content. This multifaceted tool is particularly notable for its cost-effectiveness, offering local-first operation without API keys and providing features like checkpoint resumption and speaker identification, making it an essential resource for researchers and developers working in the field of speech recognition and natural language processing.