Scene measures and content type: what they can and cannot tell you

CT-1, the first content-type model, missed. Speech was found with a recall of 0.82 and a precision of 0.56; music with 0.46 and 0.65; noise in 3 of 200 fragments (recall 0.015). Music and noise were often called speech: 55 of 200 music fragments and 75 of 200 noise fragments. The confidence barely dropped when the answer was wrong (0.76 to 0.62). On clean read speech the answer was right in 98 to 100% of cases.

CT-2, the second model, is better but did not meet its criterion. Speech recall 0.985 and precision 0.952 (first model: 0.820 and 0.558); music 0.950 and 0.864 (first model: 0.455 and 0.495); stationary noise recall 0.444 (40 of 90: town square 23 of 30, street traffic 4, metro 13) with precision 1.000. The 50 misses are all called unknown, never speech or music. It beat the first model in every class; noise was the one that failed its bar. It is an opt-in in the API, and the first model stays the default there until 15 January 2027.

MX-1: the label "mixed" is hardly used where it belongs. For speech over music 6 dB softer it was used in 29 of 100 cases (the bar was 80), and for music as loud as the speech in 36 of 100. It was used wrongly on 23 of 200 recordings that were not a mix (0.115, the bar was at most 0.10). Pure instrumental music was called music in 34 of 100 cases.

SC-1: two scene measures work, one works backwards. The share of the pauses' energy in the speech band and in the low band separates babble from machine noise: AUC 0.803 and 0.807 at 20 dB, 0.867 and 0.919 at 10 dB (with the second speech detector 0.878 and 0.901, and 0.943 and 0.968). The stationary score failed and reads the wrong way round: AUC 0.23 at 20 dB and 0.27 at 10 dB against a bar of 0.80, because dense babble scores as more stationary than a washing machine. The measure for the kind of background failed too: a machine mostly reads as unsteady.

All of these measures are Proxy: information only, never a pass or fail check. The stationary score is now called "Constant background noise", not "machine".

Content type holds for clean read speech and is a first indication for music and noise. The scene measures say something about groups of environments, not about one file.

• A television or music behind a speaker in a real room: the scene study used synthetic speech and no television or music.
• Real conversations, and other languages.
• AudioLab does not recognise machine noise, and it does not detect a second voice or a television.