Moonshine beats faster-whisper on Japanese test: lower CER, 0.2s vs 1.3s

The post is a hands-on implementation write-up aimed at developers who want to benchmark Japanese speech recognition themselves. It supplies the setup steps — a Python virtual environment with moonshine-voice and faster-whisper installed, and 16 kHz mono 16-bit WAV input produced with ffmpeg — and works with moonshine-voice 0.1.5 and faster-whisper 1.2.1. Moonshine is fed audio in 100 ms chunks at real-time speed so that each line's completion time can be recorded, while faster-whisper receives the whole clip at once after a warm-up run to avoid slow first-load times. Response speed is defined differently for each: for Moonshine, the gap between when the last line was finalized and when that line's speech ended; for Whisper, the total processing time of a clip.

The accuracy metric is character error rate (CER), computed as the edit distance against a reference sentence divided by the number of reference characters, after stripping punctuation and whitespace and normalizing full-width and half-width characters. The author flags a caveat: if "三百二十八万五千円" is written as "328万5000円", the meaning is the same but CER rises, so number formatting should be aligned before measuring. The code also sets max_tokens_per_second to 13 for Japanese, versus a default of 6.5 — a value the developer documentation recommends for languages with high tokens-per-second counts. In the author's own measurements, 6.5 and 13 produced identical CER to three decimal places, so the change made no difference on those samples, but the author suggests following the documentation anyway.

Two practical snags are documented. Feeding 20 clips through a single Python process returned empty Moonshine results for three of them, while running the same audio alone worked correctly; the cause was not identified and per-clip processes avoided the problem. And Moonshine breaks lines at breaths and pauses, so spoken Japanese comes out as many short lines that need post-processing to reassemble into sentences. The author also notes that loading a non-English model displays a non-commercial license notice, while the official repository's LICENSE lists the current streaming models, including Japanese, as MIT — a discrepancy the author advises checking officially before commercial use.