FeelFish 4.1 is officially out today.

For a while now, most of our energy has gone into the "writing" side: longer context, smarter agents, a more complete workbench. But writing doesn't only happen at a keyboard. When a line of dialogue comes to you on a walk, when you want to hear your own draft before revising it, or when a novel could become an audiobook, the keyboard becomes a wall.

4.1 is about taking that wall down. Voice input, read-aloud, and audiobook production now form one continuous chain: what you say can become text, what you write can be read back to you, and what you finish can become an audio work you can publish.

🎙️ Voice input: hold Space and talk

Voice used to be a tiny microphone button in the corner of the input box: tap to start, tap to stop. In 4.1 it becomes a mode:

A voice input toggle now sits at the top-right of the input box (default shortcut Cmd/Ctrl + Shift + D). Entering it turns the input area into a voice panel: hold Space to speak, release to recognize and insert, press Enter to send.
A quick tap on Space still types a space, so the mic never opens by accident. Press Esc while recording to cancel, or Esc when idle to exit.
Recognized text is never sent automatically. Click the text in the panel to return to the input box and keep editing, or clear it and start over.
When several input boxes are on screen, only one can be in voice mode at a time, so there are never two recordings running.

Where recognition happens matters even more. On Windows and macOS, the default is a local model: recognition runs entirely on your machine, audio never leaves your computer, no credits are used, and the model downloads on demand instead of bloating the installer. On HarmonyOS PC, cloud recognition is the default; the first time you enable it, FeelFish clearly tells you that recordings will be sent to the recognition service, and you can choose a different online model in Settings.

Voice input is a member feature. If you're not a member, the toggle shows a subscription entry rather than silently downloading a model or opening your microphone.

🔊 Read aloud: give your words a voice

Read-aloud has been rebuilt in 4.1.

A read-aloud bar now appears below the editor. Press play and the text is read paragraph by paragraph, with the current passage highlighted; you can pause, resume, jump to where it's reading, and change voice or speed at any time. For long texts, the next segment is prefetched in the background so playback doesn't stall at the end of a paragraph.

Three kinds of voice, pick what fits:

Local model: Chinese-English mixed reading, works offline after a one-time download, uses no credits, and keeps your text on your machine. Available on Windows and macOS.
Doubao Voice: cloud synthesis with a more natural delivery, uses a small number of credits, and is a member feature. You can also use your own key, which uses no credits. Both options require membership.
Online model: connect the "Speech synthesis" scenario to any OpenAI-compatible TTS service.

Changing source, voice, or speed re-reads the current segment instead of skipping ahead or waiting for the next one. Read-aloud, audio-file playback, and video preview yield to one another, so only one sound plays at a time.

If you'd like the agent's replies read out too, turn on "Auto read aloud" at the top-right of the agent panel. It's off by default; you can read every reply or only the final turn, and replies are queued so the one currently playing is never interrupted.

🎧 Audiobook production: from novel to audio work

Read-aloud is for hearing what you wrote; an audiobook is for handing your work to listeners. 4.1 adds a built-in "Audiobook production" solution that turns this into a pipeline you can inspect and roll back:

Choose the chapters to produce, and whether it's a single narrator or a narrator plus multiple character voices.
The agent turns the manuscript into a narration script and builds a cast table with each character's voice, speed, and voice profile.
Before generating, it gives you the estimated number of segments and the cost; batch generation starts only after you confirm.
Outputs are archived by chapter in the project, ready to preview or continue with the next chapter.

The work-level decisions are stored in the project: the cast table and narration style are long-lived files, so later chapters reuse the same voices and reading conventions, and only newly appearing characters need new suggestions. Want to hear a sample first? Ask for a few sample segments before committing to the whole book.

The built-in solution uses Doubao's audio generation model, so ambience, sound effects, and music can be written into the script and generated together; basic assembly doesn't require installing ffmpeg. A single audio segment is capped at 120 seconds, so scripts are split by scene, and billing is by output duration with an estimate before every batch.

The solution, skill, and production parameters can all be copied and customized — the built-in solution uses the same capabilities available to custom solutions, with no hidden switches reserved for built-ins.

Upgrade now

4.1 is out. Open FeelFish and choose "Check for Updates" in Settings, or get the installer from the download page.

If you run into a problem, or want a feature that isn't there yet, tell us anytime. We read every piece of feedback.

May your next idea arrive without waiting for you to sit back down at the keyboard. 🐟

FeelFish 4.1 Released: A Comprehensive Audio Upgrade

🎙️ Voice input: hold Space and talk

🔊 Read aloud: give your words a voice

🎧 Audiobook production: from novel to audio work

More in 4.1

Upgrade now