One of the things that bothered me while testing EmberWriter's audiobook generation was hearing a character's dialogue suddenly switch back to the narrator's voice. The words were correct, but the performance wasn't. A conversation between two characters could lose its rhythm because the system didn't always understand who was speaking.
That sent me back into the audiobook production pipeline, and it led to one of the bigger improvements we've made to EmberWriter so far.
EmberWriter RC2 introduces Audio Director, a new approach to turning novels into narrated audiobooks.
Instead of sending the original manuscript directly to the speech engine, EmberWriter now creates a separate Audiobook Performance Manuscript. It preserves the author's original writing while adding the information needed to produce a better performance: character assignments, emotional direction, delivery cues, and pronunciation instructions.
🎭 What's New in RC2?
Character-aware narration: EmberWriter identifies dialogue and assigns it to the appropriate character voice. When it cannot confidently identify who's speaking, it flags the passage for review rather than automatically giving it to the narrator.
Emotional performance direction: Authors can review and adjust cues such as fear, sadness, joy, urgency, relief, and restraint. These instructions become part of the audiobook manuscript without changing the novel itself.
Audio Director interface: Review dialogue, correct character assignments, and adjust delivery instructions directly inside EmberWriter's Audiobook Workshop.
Improved pronunciation: We've retained our CMU/IPA pronunciation improvements, including corrections for words such as buoy and buoys.
Better pacing: Chapter headings now have a deliberate 1.35-second pause before narration begins.
Smarter regeneration: Unchanged audio segments can be reused, so correcting one character's dialogue doesn't require regenerating the entire audiobook.
🎧 Hear the Results
I regenerated our original three-chapter demonstration, The Bell Below Low Water, using Kokoro ONNX.
The audiobook features five distinct voices: the narrator, Mara, Ivo, Nessa, and Sal.
Across three chapters, we identified and assigned speakers to 70 dialogue passages, including 30 additional assignments prepared for author review.
Here are three short samples from the actual recordings:
Character Voice Assignment and Performance Controls
🔥 What About Fish Audio?
Fish Speech's approach to expressive narration helped inspire this architecture.
EmberWriter now has an engine-independent performance manuscript and a Fish-compatible rehearsal workflow. The idea is to let different speech engines interpret the same performance instructions according to their capabilities.
Kokoro currently handles distinct character voices and pronunciation, but its emotional expressiveness is limited. Fish integration is still experimental, and we haven't completed live expressive-performance validation yet.
I'm particularly interested in where we can take this next: an audiobook system that understands not just what a character says, but what they've experienced in the story and how that should influence their performance.
✅ Release Candidate Status
RC2 has passed 559 backend tests, and the frontend production build completed successfully.
The new three-chapter audiobook has also passed automated checks for character assignments, pronunciation rules, and chapter pacing.
We're keeping this labeled as a release candidate while we finish listening reviews, clean-install testing, and further Fish integration.
EmberWriter remains a local-first writing and book-production environment. The goal is to give independent authors more control over the entire process, from their first chapter to a finished book and, eventually, an audiobook that genuinely sounds performed rather than simply read aloud.
If you're testing EmberWriter through Pinokio, I'd appreciate hearing how the new dialogue attribution works with your stories, particularly conversations with several characters or dialogue where the speaker isn't explicitly identified.
Those are the situations that exposed the problems in the first place, and they're exactly where I want Audio Director to keep improving.
One of the things that bothered me while testing EmberWriter's audiobook generation was hearing a character's dialogue suddenly switch back to the narrator's voice. The words were correct, but the performance wasn't. A conversation between two characters could lose its rhythm because the system didn't always understand who was speaking.
That sent me back into the audiobook production pipeline, and it led to one of the bigger improvements we've made to EmberWriter so far.
EmberWriter RC2 introduces Audio Director, a new approach to turning novels into narrated audiobooks.
Instead of sending the original manuscript directly to the speech engine, EmberWriter now creates a separate Audiobook Performance Manuscript. It preserves the author's original writing while adding the information needed to produce a better performance: character assignments, emotional direction, delivery cues, and pronunciation instructions.
🎭 What's New in RC2?
Character-aware narration: EmberWriter identifies dialogue and assigns it to the appropriate character voice. When it cannot confidently identify who's speaking, it flags the passage for review rather than automatically giving it to the narrator.
Emotional performance direction: Authors can review and adjust cues such as fear, sadness, joy, urgency, relief, and restraint. These instructions become part of the audiobook manuscript without changing the novel itself.
Audio Director interface: Review dialogue, correct character assignments, and adjust delivery instructions directly inside EmberWriter's Audiobook Workshop.
Improved pronunciation: We've retained our CMU/IPA pronunciation improvements, including corrections for words such as buoy and buoys.
Better pacing: Chapter headings now have a deliberate 1.35-second pause before narration begins.
Smarter regeneration: Unchanged audio segments can be reused, so correcting one character's dialogue doesn't require regenerating the entire audiobook.
🎧 Hear the Results
I regenerated our original three-chapter demonstration, The Bell Below Low Water, using Kokoro ONNX.
The audiobook features five distinct voices: the narrator, Mara, Ivo, Nessa, and Sal.
Across three chapters, we identified and assigned speakers to 70 dialogue passages, including 30 additional assignments prepared for author review.
Here are three short samples from the actual recordings:
🎧 Mara and Ivo — Character Dialogue (20 seconds)
https://raw.githubusercontent.com/grumpystrongman/EmberWriter-Pinokio/main/media/rc2/sample-01-mara-ivo-dialogue.mp3
🎧 The Rescue Decision (22 seconds)
https://raw.githubusercontent.com/grumpystrongman/EmberWriter-Pinokio/main/media/rc2/sample-02-rescue-decision.mp3
🎧 Multiple Character Voices (26 seconds)
https://raw.githubusercontent.com/grumpystrongman/EmberWriter-Pinokio/main/media/rc2/sample-03-multiple-character-voices.mp3
📸 Screenshots
Audio Director — Dialogue and Speaker Validation
Character Voice Assignment and Performance Controls
🔥 What About Fish Audio?
Fish Speech's approach to expressive narration helped inspire this architecture.
EmberWriter now has an engine-independent performance manuscript and a Fish-compatible rehearsal workflow. The idea is to let different speech engines interpret the same performance instructions according to their capabilities.
Kokoro currently handles distinct character voices and pronunciation, but its emotional expressiveness is limited. Fish integration is still experimental, and we haven't completed live expressive-performance validation yet.
I'm particularly interested in where we can take this next: an audiobook system that understands not just what a character says, but what they've experienced in the story and how that should influence their performance.
✅ Release Candidate Status
RC2 has passed 559 backend tests, and the frontend production build completed successfully.
The new three-chapter audiobook has also passed automated checks for character assignments, pronunciation rules, and chapter pacing.
We're keeping this labeled as a release candidate while we finish listening reviews, clean-install testing, and further Fish integration.
🚀 Install or Update Through Pinokio
New users:
https://github.com/grumpystrongman/EmberWriter-Pinokio
Existing users: Open EmberWriter in Pinokio and select Update.
Once updated, navigate to:
Publish → Audiobook → Audio Director
You can find the full RC2 update, screenshots, and audio demonstrations here:
https://github.com/grumpystrongman/EmberWriter-Pinokio/issues/1
EmberWriter remains a local-first writing and book-production environment. The goal is to give independent authors more control over the entire process, from their first chapter to a finished book and, eventually, an audiobook that genuinely sounds performed rather than simply read aloud.
If you're testing EmberWriter through Pinokio, I'd appreciate hearing how the new dialogue attribution works with your stories, particularly conversations with several characters or dialogue where the speaker isn't explicitly identified.
Those are the situations that exposed the problems in the first place, and they're exactly where I want Audio Director to keep improving.