Nanosamur.ai has added a way to run several speech-recognition models against the same live audio stream. Its demonstration compares Qwen3-ASR and faster-whisper on roughly a minute of medical dictation, showing their transcripts side by side as they change.
Qwen recorded a 2.7% word error rate in that sample, compared with 4.7% for faster-whisper. That measure counts substitutions, missing words and extra words, but does not distinguish a minor spelling error from a change to an important medical term.
In the demonstration, Qwen changed “ST elevation” to “SGL elevation.” Faster-whisper recognized that term correctly, but made other errors, including inserting “not” into a phrase. The recording also shows a pause in faster-whisper’s output.
The comparison is the platform developer’s analysis of one clip, not evidence that either model is suitable for clinical use.