Auto-Ear: 71% 'failure' verdicts caused by silence, not microphones
Analysis of 1,066 real recordings in the Auto-Ear car engine diagnostic service showed that 71% of 'failure' verdicts are not due to engine condition or phone microphones but to silence and noise in recordings. Users record 13 dB quieter than reference, and a third of recordings have over 70% silence, which the model interprets as malfunction.
The developer of Auto-Ear, a free service for preliminary engine diagnosis by sound, analyzed 1,066 production recordings from July 9 to August 9, 2026. Comparing audio features between recordings labeled 'failure' and 'normal', they found that typical 'failure' recordings are 9 dB quieter, have 43% silence versus 1%, and higher spectral flatness. Phone brand had minimal effect within the same loudness quartile, but loudness itself caused a two-fold difference in verdicts: quiet recordings are twice as likely to be judged 'failure' than loud ones. Further analysis showed that users record much quieter than reference samples (-30.6 dB vs -17.8 dB) and a third of recordings contain over 70% silence. The model, based on CLAP embeddings, had protection only for silence above 90%, passing partial silence. The developer concludes that in audio diagnostics, one should first check how the model treats absence of sound, and plans a quality filter to reject recordings with too much silence or too low volume.
- Abbreviations
- RMS = Root Mean Square — среднеквадратичное значение
- dBFS = decibels relative to full scale — децибелы относительно полной шкалы
- AUC = Area Under the Curve — площадь под кривой
Source: Habr — хаб ИИ —
original