Benchmark

OpenAMD, English subset, 128 clips: 85 people and 43 recordings answering a phone.

Accuracy and errors by model on the OpenAMD, English subset, 128 clips
Model	Accuracy	People hung up on	Recordings connected
Patch AMD	96.9%	4 / 85	0 / 43
CNN6-GBM (dataset’s best)	96.1%	4 / 85	1 / 43
CNN6-GBM-balanced	93.0%	6 / 85	3 / 43
wav2vec_finetune	92.2%	4 / 85	6 / 43
whisper-GPT-5.4	91.4%	11 / 85	0 / 43
RNN-YAMNet	88.3%	13 / 85	2 / 43
CNN6-LogReg	87.5%	15 / 85	1 / 43

Timing

People are connected a median 0.15 s after their speech ends (90th percentile about 1 s).
Every recording is hung up on before it finishes playing, a median 3.4 s after pickup.

Method

We used the OpenAMD dataset (Zenodo record 21234058), English clips only: 128 clips, 85 people and 43 recordings. Each clip was replayed in real time, as phone-quality call audio, through the same production detector the API runs, exactly as a live call would reach it. After a clip ends the line stays quiet, as when a person waits for a reply. We compared the result with the predictions the dataset publishes for its own models on the same clips. Two full runs gave identical results.

Scoring follows what a dialer does: human or not_sure connects a rep; voicemail, ivr, unavailable or silence hangs up. A person hung up on and a recording connected are both errors.

Where it missed

Two clips with almost no audible speech; they were treated as silence.
Two people who open with “you’ve reached …”, which is how most voicemail greetings start.

What this does not show

The clips are short, clean recordings in English. Real calls add line noise, crosstalk and accents; we expect real-world accuracy to be somewhat lower for every model in the table. We have not measured carrier AMD (Twilio, Telnyx premium) on this dataset, so we make no numeric claim about it — see how Patch AMD differs from carrier AMD.

Credit