Small TTS Models Post Strong Benchmarks, but Latency Claims Are Being Misreported
A blog post found that a widely repeated figure of 13.7ms time-to-first-audio for Zyphra's ZONOS3-Mini does not appear in the technical report, which lists 85ms on an H100 and 80ms at eight concurrent requests. The author argues the inflated number spread because it sounded more impressive than the paper's actual result. Several small text-to-speech models were released in mid-November 2025, including tiny-voca, Inflect v3 and sanoTTS, with the author noting that tiny-voca was trained entirely on synthetic audio generated by a larger model. Benchmark comparisons across these models are not directly matched, since they use different test panels and metric variants. The piece concludes that 'real time' performance varies by hardware and that reported figures often disagree with the underlying papers.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in