Speech recognition glossary
Short, sourced definitions of the terms used when choosing and measuring speech-to-text models.
Terms
- Word error rate (WER)
Word error rate is substitutions plus deletions plus insertions, divided by the reference word count. The formula, a worked example, and when to use CER.
- Token-and-Duration Transducer (TDT)
A Token-and-Duration Transducer predicts each token and how many audio frames it covers, so decoding skips frames. What TDT is and what its 2023 paper reports.
- On-device speech recognition
On-device speech recognition runs the speech-to-text model on your own laptop, phone or server, so audio is not sent off to be transcribed. How it compares.
- Hotword biasing (contextual biasing)
Hotword biasing tells a speech model which names and terms to expect, so it favours them when the audio is close, with no retraining. How it works, by example.
- Real-time factor (RTF) and RTFx
RTF is processing time over audio length, so lower is faster; RTFx is the inverse, so higher is faster. How to read 174x realtime, with Phonon-2's figures.
- Transducer (RNN-T) vs CTC vs attention encoder-decoder
A transducer (RNN-T) joins an audio encoder and a token predictor to emit text frame by frame. How it differs from CTC and attention models, with the papers.
- Text normalisation for speech recognition
Text normalisation rewrites reference and output into one form (case, punctuation, numbers, spelling) before WER is counted. Why scores depend on it.
- The FLEURS benchmark
FLEURS is a public Google speech dataset of read sentences in 102 languages, about 12 hours per language. What it contains and how to read FLEURS results.
- The Open ASR Leaderboard
The Open ASR Leaderboard runs speech models on the same public test sets with one scoring code. Its English sets, how to read averages and RTFx, and Phonon.
- Word timestamps in speech recognition
Word timestamps are the start and end time of each word in a transcript. How models produce them, what affects accuracy, and how to get them from Phonon.