What SNR Means for Transcription Accuracy — and the Thresholds That Matter

What is signal-to-noise ratio in audio?

SNR is the ratio between the power of the signal you care about (speech) and the power of everything else (line hiss, air conditioning, street noise, other voices), expressed in decibels. Because decibels are logarithmic, every 10 dB means a 10× power difference: at 20 dB SNR speech carries one hundred times the noise power and transcription is comfortable; at 10 dB it is ten times and errors climb; near 0 dB speech and noise are equally loud and even human listeners start failing.

Intuition by environment: a quiet office call lands around 25-35 dB SNR, a decent headset in an open-plan office 15-25 dB, a speakerphone in a busy branch or a car 5-15 dB, and a souq or roadside call can dip to single digits. Crucially, SNR is measured per call — the same agent, same headset, same engine produces different accuracy at 2 PM than during the evening rush purely because the noise floor moved.

How does low SNR affect word error rate?

Noise attacks ASR in three ways. It masks the high-frequency consonant energy that distinguishes similar words — and Arabic is especially exposed here, because phonemic contrasts like ح/هـ and س/ص live exactly in the bands that noise and narrowband telephony erase. It confuses voice activity detection, so noise gets transcribed as speech (hallucinated insertions on hold music are the classic failure) or quiet speech gets dropped as noise. And it degrades diarization, since speaker embeddings computed from noisy segments smear together and same-gender speakers stop separating cleanly.

The degradation is non-linear. From 30 dB down to about 15 dB, modern noise-robust models lose accuracy gently. Below 15 dB the curve steepens sharply, and below roughly 5-10 dB transcripts become unreliable regardless of vendor. That knee in the curve is why CallScribe states its headline accuracy claim with the condition attached — 85-95% word-level accuracy on clear audio at SNR above 15 dB — and scores every uploaded call's audio quality so a bad-audio transcript is flagged as such rather than silently trusted.

What do LUFS and speech activity add to the picture?

SNR alone does not fully describe audio quality, so production scoring adds two companions.

Together the three metrics explain most transcript-quality surprises before anyone reads a word: a call scoring 8 dB SNR, -35 LUFS, and 30% speech activity will produce a poor transcript from every engine on the market, and the correct fix is upstream in the audio chain, not a different model.

• Loudness (LUFS). Defined by ITU-R BS.1770, LUFS measures perceived loudness rather than raw amplitude. Speech sitting near -18 to -24 LUFS is healthy; a recording at -40 LUFS is so quiet that quantization noise and line hiss dominate after normalization, and clipping at the other extreme distorts the waveform itself. Telephony stacks with bad gain staging routinely record one channel hot and the other nearly silent.
• Speech activity ratio. The fraction of the recording that actually contains speech. A 10-minute call with 40% speech activity is six minutes of hold music, IVR prompts, and silence — billable minutes that produce no transcript value and, on weaker pipelines, active hallucinations. Persistently low speech-activity scores usually indicate a recording configuration problem (capturing the full hold queue) rather than a conversation problem.

Why does 8 kHz telephony audio hurt accuracy?

Classic telephone audio is narrowband: sampled at 8 kHz, it carries frequencies only up to about 3.4 kHz (ITU-T G.711, the PSTN workhorse codec). Human speech — especially fricatives like /s/, /f/, /th/ and the Arabic emphatic-plain contrasts — has significant energy above that ceiling, and the codec simply discards it. ASR models pretrained mostly on 16 kHz wideband audio then receive an upsampled signal where the top octave of information is permanently gone; no amount of model quality reconstructs distinctions the wire never carried.

In practice, expect a few WER points of penalty for narrowband audio even at good SNR, compounding with every other impairment on the call. The mitigation hierarchy is unglamorous but effective: prefer VoIP paths that negotiate wideband codecs (G.722 or Opus) end-to-end, record at the PBX or platform rather than re-compressing downstream, keep recordings in lossless or high-bitrate formats instead of low-bitrate MP3, and use dual-channel recording so each speaker's signal stays clean. Each step preserves information that every later stage — ASR, diarization, sentiment — depends on.

The operational takeaway: treat audio quality as a monitored production metric, not an assumption. Score SNR, loudness, and speech activity on every call, set alerting thresholds at the 15 dB knee, and route persistent offenders (a branch with a broken headset, a trunk with bad gain) to infrastructure fixes. Teams that do this recover more accuracy than any vendor switch would deliver.