ASR - Notes from Building Speech-to-Text

Field notes from building and tuning automatic speech recognition (ASR) pipelines - the components that sit around the model, the trade-offs that decide quality in practice, and what the experiments actually showed.

The series

• Tuning Voice Activity Detection for ASR Gating - why a VAD’s defaults aren’t enough for ASR, how tuning changes the gate, and why the right answer is often a different VAD per use case. Silero VAD and FunASR’s FSMN-VAD as worked examples.

More to come.

---

Back to the blog home.