Soniox, ElevenLabs tie on Japanese real-time transcription accuracy

The motivation for the test came from a real failure: the author's assistant transcribed a request about Kyoto's Kamogawa as two fragments, with "Kamogawa" heard as "Kanaoka." When the author looked for neutral numbers comparing ElevenLabs and Soniox on Japanese real-time transcription, none existed—Artificial Analysis covers only batch-file transcription without Japanese, and Soniox's own benchmark covers English only. That gap is why the test uses public Japanese speech read aloud, drawn from Google's FLEURS dataset with a fixed seed for 40 clips totaling 8 minutes 40 seconds of news and encyclopedia-style text full of names and numbers.

The scoring separated accuracy from boundary behavior. ElevenLabs' higher raw error rate was partly a numeral-formatting issue—it wrote "13th" and "24th" as kanji numerals where the reference uses Arabic digits—and correcting that brought its error rate close to Soniox's. The sharper split appeared in sentence segmentation: Soniox under LiveKit's plugin default split 12 of 40 clips mid-sentence, while waiting longer reduced that to a single clip, the one where a long pause after a parenthesis came through as a break. That trade-off is visible in latency tails, with Soniox's slowest tenth exceeding 2 seconds.

The author's bottom line for voice-controlled apps is to choose Soniox at a 2000 ms wait, since it cuts mid-sentence splits while keeping sentence finalization fast and costing well under ElevenLabs. The test itself is small—40 clips, roughly 9 minutes, each condition run once—so the accuracy gap should be read as a tendency, and conversational speech with fillers, dialects, or proper nouns may produce different results.