- Author
- Posted by karimf, with replies from many commenters
- Published
- 21 October 2025
- Kind
- Hacker News discussion thread about an explainer post from Kyutai
- Original
- news.ycombinator.com/item?id=45655161
- Subject
- Audio models and analysis
- Themes
- What the senses do
What it is
The thread is on a Kyutai post with the same title. It had 428 points when collected. Replies come from readers, an audio engineer, a DAW author and the post's author.
What it says
- The author says neural codecs beat MP3 and Opus on size but cost more CPU.
- Commenter bob1029 proposes feeding MP3 frames to a transformer as tokens. Replies say compressed input is hard for networks.
- One commenter reports a voice assistant misnamed a hummed tune as Beethoven's 5th.
- The post's author thinks weak tone understanding is a capability issue, not a safety one.
- The post's author says Kyutai transcribed 7 million hours of audio to train Moshi.
- Commenters dispute whether engineers can identify speech sounds from a waveform.
- The post's author doubts a physics-based speech codec will catch on.
What evidence it gives
The thread offers opinions, anecdotes and links to papers. The post's author gives training figures and compression numbers from work at Kyutai. Claims about voice models rest on informal personal tests. One commenter posts a long image-model experiment with image links.
Interests and limits
The post's author mentions Kyutai colleagues and Moshi. Those replies come from an involved party. Commenters' voice assistant tests are informal and cover products that changed over time. Claims in the thread are not checked here. The linked article was not read for this page.