Playground · ASR

Streaming speech

Operational view for diarized transcripts — word stability, WER estimates, and chunk timestamps tied to /v1/audio/transcribe.

Model: dex-whisper-edge

Sample rate: 48 kHz

WER (eval)

4.2%

RTF

0.31

Partial rate

218/s

GPU util

62%

Waveform · session s04

00:01

Establishing QUIC stream to regional POP.

00:03

VAD locked · language auto: en-US.

LIVE

Final transcript · diarized (mock)

[spk_0] 0.0–4.1s
“We need the latency envelope under forty milliseconds before the board review.”

[spk_1] 4.2–7.8s
“Copy — I’ll pin dex-whisper-edge to the eu-west shard and replay the golden set.”

[spk_0] 7.9–11.2s
“Good. Emit word-level confidence so support can filter low-trust tokens in the UI.”