Forge & Voice

Built from a love of audiobooks

I love books, but I’m often walking or busy doing something else. Audiobooks let me keep following a story when I can’t sit down and read. That is the personal reason behind Forge & Voice: I wanted to explore turning written books into audio I could enjoy while going about my day.

That idea led me to explore a local audiobook pipeline and then build a studio for reviewing text, choosing voices, and producing recordings. Giving characters distinct voices became part of making the listening experience more expressive. The current studio works with shorter passages and scripts; extending it to long-form chapters is still a next step.

The first recording was only the beginning

Generating speech exposed the work around the model. A saved voice sample could be mistaken for a reading of new text. An unavailable speech service could leave someone waiting without knowing why. In a cast recording, one loud reference could also make a character dominate the conversation.

Those problems shaped the interface: reference previews and generated recordings now have separate controls, service readiness is visible, and recording levels are checked and balanced. Users can audition a voice before committing to a longer production.

Keeping the writer in control

I built the script workflow around a review step. Imported text becomes structured JSON, the user checks the spoken words, and each speaker gets a chosen voice. The runner preserves the source through segmentation and stops an incomplete set of audio segments from becoming a finished export. A saved production record keeps the cast, script, and output together after a restart.

What the studio became

The current app supports both a quick single-voice passage and a reviewed multi-speaker production. A local MCP interface exposes the same workflow to an assistant. In the latest end-to-end check, a two-speaker script produced a downloadable recording.

The project taught me to treat previewing, review, failure handling, and recovery as part of audio production. It uses reference-based voice cloning, not newly trained speech weights. Long-form chapter editing and verification of the actual spoken words are the next steps.

From a document to a cast recording

The studio imports TXT, Markdown, DOCX paragraph text, and its own JSON format. It recognizes explicit speaker labels, preserves spoken words, and asks the user to review the structured script. Each speaker gets a chosen saved voice before generation. The production manifest records the cast and each audio segment, making the result easier to inspect and reproduce.

A local Model Context Protocol server exposes the same script conversion, validation, cast generation, voice listing, and job-status actions to an assistant. It has been tested over stdio and registered in Codex. A key lesson was separating three different operations: preparing a voice reference, assigning that reference to a character, and training new model weights. This version implements the first two and provides a training-readiness report; it does not claim a new fine-tuned model.

Make a voice, then audition it

Add a clear 10–30 second recording and its exact transcript. Preview the upload, confirm permission, and prepare a local voice profile. The quality panel reports duration, recording level, and possible clipping. Generate a short test with new words and compare it with the source recording.

Existing profiles can be revised as separate copies: correct the transcript or adjust recording level, then audition both versions. This uses Fish reference-based cloning; it does not fine-tune model weights. Signal checks catch some recording problems but do not establish speaker similarity or pronunciation quality.