Reka Speech High Throughput Speech Transcription And Translation Model With Timestamps
Source
reka.ai
Author
Reka AI
Date
Context
Reka AI designed Reka Speech to produce transcripts with timestamps efficiently enough for large audio archives. During transcription, its system saves internal attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → inputs in CPU memory. It later uses them to reconstruct attention scores and align the generated words with the audio. Those timestamps make the text usable for subtitles and locating passages in recordings. In a test of 16,390 English audio files on one H100 GPU, Reka reported processing the dataset 8 to 35 times faster than three Whisper-Medium-based alternatives. That comparison concerns the tested transcription-and-timestamp pipeline. Technical account.
Terms in this piece · Glossary
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
