Audio Meter Moving but Words Wrong? What to Check | Yihu
Replay saved audio, check names, numbers and negation, then review input and transcription separately. Try Yihu’s no-mic audio checklist and review privacy before installing.
Preview privacy before you try
The linked preview uses preset material, needs no login or microphone, and keeps edits only in the open page. Reloading resets edits. Optional read-aloud uses the browser’s voice service; usage analytics require consent.
Keep private originals in approved tools. In the app, on-device recognition and the destination’s storage, sync and access rules are separate steps. Review the installed version before using a source-described workflow. Read the data-flow guide
US listing verified; availability depends on your storefront. No Mac or self-signing for store installation. Bring your workspace address and account; the optional speech model is about 950 MB.
Replay saved audio, then compare the words and input
Keep saved audio and text first. Replay a short section and check names, numbers and negation. If the words differ, review the actual input, speech model and final transcript separately; do not raise input merely to fill the meter. Short speech and long notes use different RMS scaling. Meter levels are not recognition confidence; compare the transcript with the saved audio.
Different from silence, voiceprint and original-word review
Diagnosis checks missing audio, voiceprint review checks filtering scope, listening checks content. Here what meter formulas/displays establish matters. No real capture or system-volume change occurs.
| Check | Verified scope | Manual review | Not established |
|---|---|---|---|
| Short speech | RMS multiplied by 2, bounded | Display scale/raw samples | 0.5 means 50% recognition confidence |
| Long note | RMS multiplied by 12, capped at 1 | Different path scaling | Same length means same pressure/quality |
| Clipping | Many large values look full | Signal versus display saturation | Full means every word correct |
| Near zero | Can be quiet/no new samples etc. | Actual state/input/errors | Permission absent or nothing captured |
| Timer display | Reads current pipeline level | Display time versus final | Zero latency/all audio saved |
| RMS | Sample-energy statistic | No device/environment calibration | Standard dB SPL/noise certification |
| Text/identity | Separate ASR/voiceprint paths | Actual results/originals | Energy proves recognition/speaker match |
| Handoff | Target/receipt separately checked | Saved versus delivered | Screenshot proves business completion |
Synthetic arithmetic
Two synthetic samples 0.1 and -0.1 have RMS 0.1. The short-speech illustration is about 0.2; long-note scaling gives 1.2 capped to 1. This is math, not microphone measurement, physical pressure or comparative performance.
Inspect real content/errors even with a full meter. Display does not establish permission, model readiness or selected target. Public checks need no customer recordings; no device/wind/accuracy test ran.
FAQ
Does full scale mean 100% recognition?
No. Display clipping is not ASR confidence.
Can equal bars compare path quality?
Not directly; scaling differs and no acoustic calibration is supplied.
Does zero prove no permission?
Not independently. Review permission, route, capture and errors.
Is RMS a standard decibel reading?
No calibrated dB SPL measurement/certification is provided here.
Does a normal meter prove identity or complete content?
No. Filtering, original words, saving and transcription differ.
Were noise/microphone tests run?
No. Synthetic arithmetic is not actual device/scenario evidence.
Reading and source
Input diagnosis · Voiceprint scope · Audio/text evidence · Preview/final · Start · Privacy
