Typhoon ASR Streaming
Steerable Low-Latency Thai ASR

Choose a voice sample

Each clip is a real utterance from the paper's evaluation data — three TVSpeech broadcast clips packed with code-switched English jargon, and two GigaSpeech2 clips where the hard case is a rare Thai person name. Selecting one feeds the same audio to every model at once.

e.g.The Agri-industry clip says “ลพบุรีมันจะมีโรงงานของ CP มันจะมีโรงงานของ Betagro” — without steering the model hears the brand as “BBTโก”.

Tune the decoder live

Three independent controls let you steer accuracy vs. speed on the fly. Changing latency re-streams the clip; toggling fusion or adding a term updates the output in place — no re-run needed.

e.g.Chunk latency sets how much audio the model waits for (lower = snappier, harder). Fusion biases decoding toward a keyword list, and boost terms inject a specific word — e.g. add “Betagro” to fix the hard term in this clip.

Transcribe your microphone

A browser microphone stream now feeds a Python WebSocket server that follows the OpenAI Realtime event pattern. The same Typhoon streaming decoder handles partial hypotheses while preserving the latency and steering controls.

Watch the models transcribe

All three decoders stream the clip side by side so you can see the trade-offs directly. Highlights mark how each token was recognised, and the strip below every card reports the live metrics that back the paper's claims.

e.g.Drop latency to 80 ms: the full-context baseline collapses into garble, while the cache-aware models keep a low CER. Metrics — CER (error), recall (keywords), RTF (speed) — update as you steer.