Transcription quality
Word confidence, speaker attribution and acoustic line quality: how VoiceLens tells you how much to trust a transcript, and what each signal means.
Every finding on a scorecard is only as good as the transcript underneath it. Rather than hide that, VoiceLens grades the transcript itself and reports three separate reliability signals. They never lower the health score, because a noisy line is not the agent's fault, but they tell you when a verdict deserves a second look.
Transcription
Recordings are transcribed with word-level timestamps, per-word confidence, speaker labels and (for English) per-turn sentiment. Stereo recordings with one party per channel are transcribed multichannel, which gives exact speaker attribution; mono recordings are diarized, which infers speakers from the audio and carries its own confidence.
Supported input formats are mp3, wav, m4a, ogg/oga, webm and flac. The language can be fixed on the channel, fixed on a single upload (which overrides the channel), or left to automatic detection. Either way it is a BCP-47 tag such as en, de-DE or sl.
Word confidence
Each word carries a confidence between 0 and 1. The scorecard aggregates them into:
- Average word confidence for the call.
- Low-confidence word count and a sample of the affected words (words below 0.7 are shaded in the transcript).
When words fall below 0.5, VoiceLens raises an ASR_LOW_CONF info fault whose evidence lists up to five examples and the share of the call affected. Slot values extracted from low-confidence regions are still shown, but this is the flag that tells a reviewer to listen before trusting a captured phone number.
Speaker attribution
For diarized (mono) audio, each word also carries a speaker-attribution confidence. When it falls below 0.15, a DIARIZATION_UNCERTAIN info fault is raised: the transcript may be attributing a sentence to the wrong party. This matters most for compliance verdicts, where "who said it" is the whole question.
The scorecard states whether attribution came from multichannel audio or from diarization, and shows the average speaker confidence. If your platform can record the two parties on separate channels, do it; it removes this uncertainty entirely.
Line quality
Independently of the transcript, an acoustic pass analyses the audio itself (resampled to 16 kHz) and produces a line quality score from 0 to 100, with 80 and above considered clean. The score is built from five measurements, each of which can raise its own LINE_QUALITY info fault with the time ranges affected:
| Signal | Default threshold | What it usually means |
|---|---|---|
| Signal-to-noise ratio | below 15 dB | Background noise, speakerphone, a poor headset |
| Mains hum | at or above -10 dB | Electrical interference at 50 or 60 Hz |
| Hiss (spectral flatness) | at or above 0.4 | Broadband noise from a bad line or codec |
| Clipping ratio | at or above 0.0005 | Input gain too high, distorted speech |
| Dropouts | detected gaps in the signal | Packet loss on VoIP, network trouble |
Per-channel results are reported separately for stereo recordings, so you can tell whether the problem was on the agent's side or the caller's.
The call summary exposes lineQualityScore and snrDb, so channel analytics can show whether a specific phone line or carrier consistently produces poor audio.
How to use these signals
- Treat
ASR_LOW_CONFandDIARIZATION_UNCERTAINas review triggers for compliance-sensitive calls: verify the phrase verdicts against the audio before acting on them. - A channel whose line quality is consistently low is a telephony problem, not a coaching problem. Fix the headset, the codec or the trunk.
- When calibrating a rubric, exclude calls with very low line quality from the golden set; they measure the line, not the rubric.
- Prefer stereo, one-party-per-channel recordings wherever your platform allows it.