A successful transcript can still be a bug.
Two issues, two different fixes
The retry reproduced a PyAV decoder-argument incompatibility. The adapter validated PCM16 WAV frames directly and passed normalised samples to Whisper. The earlier category-test startup timeout was not reproduced, so its original cause remains unconfirmed.
A three-second silent recording initially produced “Thank you.” Enabling voice-activity filtering made that same fixture return an empty transcript. One silence success does not eliminate hallucinations on noise or other non-speech inputs.
Names deserve explicit tests
The collection message transcribed exactly, while the longer policy message consistently rendered “Kampala” as “compiler”. Synthetic English is useful for adapter checks, but cannot validate local pronunciation, accents or language support.
The complete request path
Both a synchronous gateway request and a queued job completed, with recorded audio-second usage matching the configured rounding rule. The rate card was synthetic. Later real-recording and mixed-load studies extended this early feasibility check.
Results, in context.
| Fixture | Audio (s) | Run 1 (s) | Run 2 (s) |
|---|---|---|---|
| Collection | 4.12 | 0.563 | 0.562 |
| Payment | 6.34 | 0.681 | 0.678 |
| Policy | 18.71 | 1.362 | 1.355 |
Separate cold request: 0.812 seconds. Timings exclude public-network transit.
How we ran it.
- Cape Town T4; pinned Faster-Whisper large-v3, Faster Whisper 1.2.0, CTranslate2 4.8.2, CUDA INT8_FLOAT16, beam five, English hint and VAD.
- Three eSpeak NG synthetic English clips at 145 words/minute, converted to mono 16 kHz PCM16, repeated twice after a separate first request. Instance-loopback HTTP.
The limits matter.
- Three synthetic clips are not an ASR quality benchmark. No real accents, Swahili, Luganda, long recordings or concurrent media traffic were tested here.
- The test proves behavior for one silence recording, not universal hallucination prevention.
Evidence & sources.
The public extract contains the methods, selected results and caveats in this note. It is a curated summary, not the full raw run. Fingerprints identify the retained source artifacts; gated audio and private operational records are not included.
Download results extractSource artifact fingerprints · SHA-256
0bbea356a922eb6d8af401e93b654988caa0e8e391806c003bc5b7ee8ace227d36570049f018a7dd0801224fce59e19e7c7b0b6b019c96bb5d1b88f3ceb3e59aSuggested citation: Embiro Labs Research. “Silence should stay silent.” Published 12 Oct 2026. embiro.ai/research/whisper-silence-and-decoding
Report a correction
