murmur: an always-on companion radio for the terminal, voiced by fish-speech

[wine-fall]

Sharing a project that uses fish-speech as its voice, in case it is useful to others here.

murmur is an always-on AI radio host for the terminal: it picks its own topics and talks, plays music, comes back and keeps going; you type a reply on the keyboard and it answers in a voice that sounds human. One host whose character holds and whose memory of you grows.

• Repo: https://github.com/wine-fall/murmur (MIT, TypeScript)
• Install: npm install -g murmur-radio && murmur (Node >= 24)
• Landing page with a two-and-a-half-minute recording of it running: https://wine-fall.github.io/murmur/

How fish-speech is used:

• The voice is fish-speech reached over a hosted endpoint (the user points murmur at an endpoint in ~/.murmur/voice.json); murmur ships no model and does not redistribute weights.
• Speech is produced ahead of time for the next program segment so there is no dead air, then mixed over the music in a Web Audio graph with gain ducking: the host talks over the song and an interjection ducks the music rather than stopping it.
• Keyboard in, voice out, on purpose: there is no ASR, all the effort went into making the output sound like a person on the radio.

Thanks for fish-speech; it is what makes the "sounds human" part true. Feedback on the voice welcome.