How LyricWinter Works: Every Character Gets Their Own AI Voice

Readers use LyricWinter to listen to fanfiction with a different voice for every character, or to save a chapter as an MP3 for a car ride. Writers use it to hear their own chapter read aloud before they post it. It works best in scenes with lots of dialogue. With one voice, you guess who is talking. With a cast, you know at once.

Hear it: a full cast, then one voice

"Floor Nineteen" is an original scene written for this blog. A maid of honor is stuck in a hotel elevator with her ex, and the night clerk keeps cutting in over the intercom. Studio found the four speakers (Narrator, Tamsin, Dario and Bev), cast a voice for each, and set how each line should sound.

Now the same scene with one voice for every character, which is how single-voice read-aloud sounds.

Listen to the four lines from "Eleven months, Dario." to "It was a very nice toast!" None of them has a "she said" or "he said." With one voice, you work out the speaker from the words. With the cast, you hear the switch. Priya, the sister in Tamsin's toast, is mentioned but never speaks, so she isn't in the cast.

How does LyricWinter work? From story to cast in four steps

You bring the story. Start a new Studio project, and Studio asks you to sign in with Google or Apple. Then pick one way in:

• A whole AO3 work, one AO3 chapter or an EPUB ebook: in the New project window, choose Import and pick AO3 work, AO3 chapter or EPUB file. Then paste the AO3 link or choose the EPUB file.
• Your own text, a file or another link: type a name under Empty project and choose Create. Add a New section with the + button at the top. Paste your story into the empty page, or choose Import story in the three-dot menu to bring in a file or link.
• No story yet? In an empty section, choose Create with AI in the three-dot menu. It writes one for 0 words.

When you choose Create audio at the top of the editor, Studio does four things.

1. It finds every character

An AI reads your story and lists everyone who speaks. It merges different names for the same person ("Sherlock" and "Holmes" become one character) and adds a Narrator for everything outside the quotes. Characters who are only mentioned are left out.

To meet the cast before you spend any words, choose Manual voice assignment. It sits next to Create audio. On a phone, it is in the small arrow menu on Create audio. Despite the name, it finds the characters and picks their voices for you, for 0 words. Then you can check and change any voice yourself. When you are happy, choose Create audio. If you skip straight to Create audio, it finds the characters and picks voices first.

This matters most in hard scenes. "Fog on the Marigold", another scene written for this blog, has two women at a ship's rail. Both are "she" for several lines, and one reply has no "she said" at all. In that recording, Studio gave every line to the right speaker. You can hear the Marigold scene and test your own dialogue.

2. It splits narration from dialogue

Next, the AI splits the story into lines and marks who says each one. It is told to keep every word in order and not to reword anything. Words like "Dario said" stay with the narrator. If a line lands on the wrong speaker, you can change it (see Direct and fix any line below). This paragraph from the scene:

Tamsin stared at the dark panel. "No. Absolutely not. Not tonight."

became two lines in the full-cast sample:

Narrator: Tamsin stared at the dark panel.
Tamsin:   "No. Absolutely not. Not tonight."

3. It casts a voice for each character

LyricWinter looks for a voice in this order:

1. Voices you pinned. Pin a voice to a character name once, and it is used whenever that name appears in a new project.
2. Voices you can use. Your uploaded and designed voices, voices friends shared with you privately, OpenAI's 13 ready-made (preset) voices, and a small library of community-created voices. The AI picks the best fit for each character.
3. A new voice built for your story. If nothing fits a character, LyricWinter designs a new voice from their own lines, for 0 words. It is private to you and yours to reuse. The narrator always gets an existing voice.

Each voice then gets an engine. Studio starts new casts on ElevenLabs v4 whenever the voice's sample recording is ready. You can move any character to another engine before you create audio.

4. It plans how each line sounds, then makes the audio

Each line becomes its own audio clip. An engine only sees the line it is given, not the scene around it. So LyricWinter reads the lines around it and adds a note on how to say it. Each engine takes that note in its own way:

• ElevenLabs v4: an audio tag, a word in square brackets such as [whispers].
• Inworld TTS 2 and OpenAI GPT-4o mini TTS: a short written instruction.
• Fish Audio S2 Pro and S2.1 Pro: an instruction added to the text the engine receives.
• Cartesia Sonic 3.5: an emotion setting chosen from the scene.

This is Auto emotion, in the small arrow menu on the Create audio button. It is on by default and only adds notes to lines that don't have one yet. On ElevenLabs v4, the added tags count toward your words, and Studio shows about how many before you start.

Studio makes many lines at once, and you can start listening before the whole section is done. To direct lines by hand, see the guides to ElevenLabs v4 audio tags, Inworld TTS steering, Fish Audio emotion tags and Cartesia emotion controls.

A different voice engine for every character

Studio picks an engine for each voice, and you can change it for any character. Keep your main characters on ElevenLabs v4, put the narrator on Inworld TTS 2, and try a side character on Cartesia Sonic 3.5, all in the same story. You don't need an account with any of these companies. LyricWinter runs the engines and charges everything in LyricWinter words.

Clone once, use it on up to 9 engines

A voice you upload works on up to 9 engines. The other 3 are OpenAI's, which use only OpenAI's own preset voices. Once the voice's sample is ready, LyricWinter prepares each engine's version the first time you use it, so there is nothing extra to upload.

Not sure which engine suits a voice? Open the voice in My Voices and choose Compare voice engines. It plays the same test line on each engine the voice can use. Choose Use on the one you like to make that engine the voice's default. Each new sample costs the words in one short test line, and samples you already made replay for free.

Mixing engines in one scene

This version of "Floor Nineteen" keeps Tamsin and Dario on ElevenLabs v4 and moves the Narrator and Bev to Inworld TTS 2.

The Narrator and Bev say about 180 of the scene's 300 words. Once v4 is 3×, moving those two to Inworld TTS 2 (1×) makes the scene cost about 40% less than an all-v4 cast. This leaves out the extra words that audio tags add. The more narration your story has, the more you save with the narrator on a 1× engine.

To switch an engine, choose the people button at the top of the editor to open the Characters tab. Pick a new engine from the menu next to the character's voice, then confirm.

Voices that stay with your characters

• Pin a voice once. In the Characters tab, open a character and choose Pin as default. That character then sounds the same in every future project. Pins work by exact character name. See how pinned voices work.
• Cast a whole project once. Characters belong to the project, so every chapter shares the same cast and voices.
• Design a voice from words. Describe the accent, age, tone and personality in Studio's Voice direction field, or let LyricWinter design one from a character's dialogue. Designing costs 0 words, up to 100 voices a day.
• Clone a voice. Upload or record a clean clip of one speaker, 10 to 30 seconds long. You confirm you have the rights to the voice (permission to use it), and the voice is always private. Our guide to voice sample length explains what makes a good clip.
• Share a voice with a friend. Invite someone by email to use a private voice. It stays private to the two of you.

Direct and fix any line

Automatic casting gives you a strong first draft. When a line needs work, you fix that line, not the whole story. Each line in the editor is called a block. Many of the tools below are in the three-dot menu at the top of the editor.

• New takes. Make a New take (a fresh recording) of any line. Older takes stay in Audio versions, and you can switch back to any of them.
• Who says what. Choose a line's speaker name to open Block settings. There you can change the Speaker or the Block type, such as Dialogue or Narration. To cut a line in two or join it to the one before, use Split at cursor or Merge with previous block in the three-dot menu.
• Controls for each engine. Block settings also show the controls the line's engine has. Inworld TTS 2 takes a written note on how to speak, a speaking speed, and a Stable, Balanced or Creative mode. Cartesia Sonic 3.5 has 58 emotion styles plus speed and volume. Fish Audio has speed, volume and expressiveness (how much the voice varies). On ElevenLabs v4, you type audio tags in square brackets right in the text.
• Plan the emotion first. Plan Emotion Steering suggests how each line in a section should sound, up to 100 lines at a time. You can review the suggestions before you create audio.
• See what the engine hears. Text sent to the voice model shows the exact text a line sends to its engine, with pronunciation fixes and directions added.
• Exact pauses. Insert silence adds a pause of 1 millisecond to 5 minutes at the cursor. Studio inserts real silence, so it works the same on every engine.

Fix pronunciation across three engines

Engines often guess at made-up names, like the place "Ceirnach Sound" in the Marigold scene. First, make the section's audio. Then put your cursor on the word and choose Fix pronunciation in the three-dot menu. If you like, type how it should sound. Studio makes six samples: Inworld TTS 2, Cartesia Sonic 3.5 and Fish Audio S2.1 Pro each read it as written and with your new pronunciation. Studio shows the cost before it starts.

Pick the one that sounds right. Change affected blocks fixes just the lines with that word, and can also move those lines to the engine you picked. Change speaker-wide fixes it for that character.

Free AI sound effects and your own sounds

Turn on AI sound effects in the small arrow menu on the Create audio button. They add sounds like the elevator and intercom in "Floor Nineteen", and they cost 0 words. You can also upload your own sounds and background ambience, up to 65 minutes per file. Turn on Use my uploaded sounds in the same menu, and Studio places them for you.

Share it, embed it, or export it

• Listening links. In the audio player's three-dot menu, choose Share listening link, then Create listening link. You get one link that stays current as you edit. Listeners see the full text with speaker labels, the line playing is highlighted, and they can play from any line. Links work for one section or a whole project, and Stop sharing turns a link off right away.
• A player for your website. For a section, the same window has Prepare audio embed and Copy embed code. They give you an audio player to paste into your website, with a "Listen and read on LyricWinter" link back to the read-along page. It always plays your latest audio.
• Files. Export final mix, in the same three-dot menu, saves a section as MP3 (192 kbps), WAV or SRT subtitles. Every line's audio must be up to date first.

You own the audio you make, so you can use it however you like, as long as you have the rights to the voices and the text.^[1]

Use LyricWinter from Claude, ChatGPT or your own code

This part is for people who use AI assistants or write code. MCP (Model Context Protocol) is an open standard that lets AI assistants connect to apps.^[2] LyricWinter runs an MCP server at https://lyricwinter.com/mcp. Add it as a connector (an add-on) in Claude or ChatGPT, and sign in with your LyricWinter account. Then ask the assistant to turn your text into a multi-voice audio story. It can create the story, cast the voices, make the audio, and hand back audio you can play.^[3]

Developers can use the REST API at https://lyricwinter.com/api/v1 to run LyricWinter from their own code. It uses scoped API keys (keys limited to the tasks you allow), made under Developers then API Keys in your account dashboard. It covers Studio projects, sections, speakers, audio workflows, playback, exports and sharing.^[4] The MCP server and the API are both in beta, and setup is in the LyricWinter API and MCP docs.

Long fics and books: one project, one cast

For a whole fic or book, start a new project, choose Import, and pick AO3 work or EPUB file. Each chapter becomes a section, up to 500 chapters. You review the list first, and nothing is created until you confirm. Every section shares the project's cast, so a character keeps their voice from the first chapter to the last.

To add a chapter later, choose New section, then paste the text, or choose Import story and bring it in From file or From link. For AO3, pick Current chapter or Full work. In a section that already has text, Append adds to the end and Replace starts the section over. A section holds up to 30,000 characters of text on the free plan, 90,000 on MyWinter (the $6 plan) and 150,000 on UltraWinter. On a paid plan, Divide into sections splits a longer story into sections that fit.

For a long fic chapter by chapter, our guide to turning an AO3 fic into an MP3 walks through the whole process. New to listening? Start with how to listen to fanfiction and our guide to AO3 text to speech. Comparing tools? See the best text to speech for fanfiction, the best AI audiobook generator for multi-voice audio and LyricWinter vs ElevenLabs for story audio.

How much does LyricWinter cost?

You pay for audio in words, which work like coins. Finding the cast is free. On a 1× engine, each word of your story uses about 1 word of your balance. The free plan refills to 2,500 words on the first of each month. That is enough for about eight scenes the length of "Floor Nineteen" on a 1× engine.

After that, 2,500 words cost $1, on the $6 monthly plan or in word packs you buy once. So a 10,000-word chapter costs about $4 on a 1× engine. On ElevenLabs v4 from October 12, 2026 (UTC), it costs about $12, plus the words Auto emotion adds. Current plans are on the pricing page.^[5]

Words you buy don't expire, and Studio charges only for lines that finish.

Is LyricWinter right for you?

LyricWinter is for you if you want to hear a specific story, like a fic you love or a chapter you wrote. Every character gets their own voice, and you choose the engine behind each one. Good to know before you start:

• Links work from four story sites: AO3, FanFiction.net, SpaceBattles and Substack. Direct links to documents work too. For anything else, including AO3 works locked to registered users, paste the text or open a file.
• It is built for English. There is no language setting for narration.
• LyricWinter Studio is in beta, so some screens and labels may change.
• Studio exports one file per section, usually a chapter, rather than one file for a whole book.

Listening to someone else's fic? Read the author's notes on the original page first, since AO3 imports leave them out. Keep the audio to yourself, and leave kudos.

The quickest way to judge LyricWinter is to hear your own story. Start a Studio project, bring in a chapter, and choose Manual voice assignment to see the full cast before you spend a word. For a whole fic or book, choose Import to bring in every chapter and cast them once.

How LyricWinter works: FAQ