IndexTTS2 Workflow Studio

Create natural, multi-speaker conversations with IndexTTS2 - fast.

Turn scripts into realistic dialogue with speaker prep, line-by-line review, regeneration, timeline editing, and polished local export on top of the official IndexTTS2 models.

Bring your own voices. This repo does not ship bundled voice clones or private speaker files.

Start In 2 Minutes

If you already have Docker and an NVIDIA-ready setup, the fastest path is:

1. Put your models in shared/models/checkpoints
2. Run docker\start.bat
3. Open http://localhost:3000
4. Start in Speaker Prep, then move through Conversation Workflow, Conversation Results, and Timeline Editor

Good entry points:

• Demo scenes: Hear What It Can Do
• Quick start: Quick Start
• Manual with screenshots: docs/manual/USER_MANUAL.md
• Docker details: docker/README.md
• Contributing: CONTRIBUTING.md
• Security reporting: SECURITY.md

Hear What It Can Do

If you want proof before setup details, start with three short scene demos rendered through this workflow:

Demo	Play	What it shows
Podcast roundtable	Play in browser	Quick multi-speaker turn-taking and pacing
Audiobook night train	Play in browser	Steadier narration with longer phrasing
Game dialogue checkpoint breach	Play in browser	Tighter back-and-forth with more urgent timing

These are public-safe sample scenes rendered from the local speaker library already used in the app.

Workflow At A Glance

flowchart LR
    A["Speaker Prep<br/>clean and score source clips"] --> B["Conversation Workflow<br/>generate first passes"]
    B --> C["Conversation Results<br/>compare takes and regenerate weak lines"]
    C --> D["Timeline Editor<br/>arrange timing, overlaps, and scenes"]
    D --> E["Export Mix<br/>render the final scene"]

Loading

For a moving walkthrough, open the timeline workflow clip.

What This Repo Is

This repository is not the official IndexTTS model repository.

It is a practical local app built on top of the official IndexTTS ecosystem, with:

• Speaker Prep for cleaning and evaluating source clips
• Conversation Workflow for fast multi-speaker script generation
• Conversation Results for review, version selection, and regeneration
• Timeline Editor for scene timing, overlaps, and final arrangement
• a Docker-first local runtime with GPU-first behavior and CPU fallback

If you want the upstream project, papers, and hosted demos, use:

• Official code: index-tts/index-tts
• Official model: IndexTeam/IndexTTS-2
• Official demo: IndexTTS-2 Demo
• Paper: arXiv 2506.21619

Why People Use It

This app is built for longer-form, practical TTS work rather than one-off single-line demos.

Example use cases:

• build a 3-speaker podcast or panel conversation
• create character dialogue for games, machinima, or visual novels
• generate narration with multiple takes, then compare and regenerate weak lines
• clean source clips before cloning so the voices sound more stable
• arrange interruptions, overlaps, and scene timing in a timeline before export

Why This Workflow Is Different

What you want to do	Where you do it	Why it matters
Clean bad source audio before cloning	Speaker Prep	Better prompt clips usually mean better voice match, less noise, and fewer robotic takes
Turn a script into multi-speaker dialogue quickly	Conversation Workflow	Fastest path from script to first usable versions
Keep only the lines that actually sound right	Conversation Results	Review, compare, edit, regenerate, and lock final takes before export
Shape interruptions, timing, and full scenes	Timeline Editor	Useful when a plain linear conversation is not enough

Main Workspaces

1. Speaker Prep

• upload or select raw source clips
• trim, convert to mono, normalize, and clean noisy audio
• run quick clone-readiness diagnostics
• save the improved result into the live speaker library

2. Conversation Workflow

• paste a multi-speaker script
• see available voices clearly
• apply pacing presets
• parse, generate, and save project state

3. Conversation Results

• compare versions line by line
• play clips, compare clips, and review scores
• edit text during review
• regenerate weak lines
• export only after every line has a chosen final take

4. Timeline Editor

• build a scene directly in the timeline
• add speaker tracks and segments
• move segments in time
• shape overlaps and interruptions
• preview and export the final arranged scene

Studio Preview

Speaker Prep	Conversation Workflow

Conversation Results	Timeline Editor

Quick Start

Recommended: Docker

This is the supported runtime path for this repo.

Default behavior:

• use your NVIDIA GPU when Docker can access it
• fall back to CPU only when GPU runtime is unavailable

1. Put your model files in:

shared/models/checkpoints

2. Start the app:

docker\start.bat

Or manually:

docker compose -f docker/docker-compose.yml -f docker/docker-compose.gpu.yml up -d --build

3. Open:

• Frontend UI: http://localhost:3000
• Backend API: http://localhost:8001
• API docs: http://localhost:8001/docs

4. Add your own clips:

• finished cloning prompts -> shared/audio/speakers/
• raw clips for prep -> shared/audio/source_clips/

5. Work through the app in this order:

• Speaker Prep
• Conversation Workflow
• Conversation Results
• Timeline Editor

To stop the stack:

docker\stop.bat

If shared/models/checkpoints is empty, the backend can automatically download the official IndexTTS2 model bundle on first start.

Official Model Downloads

If you prefer to download the models yourself instead of using the app's automatic bootstrap, use the official upstream links:

• IndexTTS-2 on Hugging Face: IndexTeam/IndexTTS-2
• IndexTTS-2 on ModelScope: IndexTTS-2

For this app, place the downloaded files in:

shared/models/checkpoints

The app is built around IndexTTS2. Older upstream model releases exist, but this repo's current workflow and docs are centered on the v2 model line.

What You Get Out Of The Box

• Docker-first startup with GPU-first behavior and CPU fallback
• DeepSpeed-enabled local inference path
• Speaker prep and diagnostics before cloning
• Multi-speaker script generation
• Review, regeneration, and final selection gating
• Timeline-based arrangement and scene export
• Project save/load for longer sessions
• Public demo scenes, screenshots, and short walkthrough videos
• Optional WebMCP website bridge for AI agents

Community And Support

• Start with the pinned support issue for common first-run help and runtime fixes
• Use the built-in GitHub issue templates for bug reports, setup help, and feature requests
• Read CONTRIBUTING.md if you want to send changes back
• Read SECURITY.md for responsible vulnerability reporting

AI / Agent Access

The web UI remains the main interface for human users.

For agent-style use, the frontend now includes an optional WebMCP bridge that can expose studio tools, prompts, and resources from the running website when the WebMCP script is available.

This is intended as a lightweight AI integration layer on top of the existing web app and API, not as a replacement for the normal UI.

Hardware And Runtime Guidance

Practical guidance for a good local experience:

• NVIDIA GPU is strongly recommended
• CPU fallback works, but startup and generation are much slower
• at least 16 GB RAM, 32 GB recommended
• allow 50 GB+ disk space for models, caches, and outputs
• the first DeepSpeed-enabled startup can take longer while extensions warm up or compile

Today the Docker image is NVIDIA/CUDA-based. AMD and Apple GPU paths are not supported by this Docker image.

More runtime details: docker/README.md

Bring Your Own Voices

This app is intentionally BYO voice.

It does not include celebrity voices, personal voice libraries, or redistributable speaker packs.

Expected folders:

• shared/audio/speakers/ - live speaker prompt files used by the app
• shared/audio/source_clips/ - raw clips for cleanup and preparation
• shared/audio/speakers_backups/ - backups of original speaker files before replacement

Quality Tips

If a voice sounds too fast, robotic, or less faithful than expected:

• use the Clone Fidelity preset in the UI
• keep random sampling off
• use natural punctuation and sentence casing in the script
• prefer a clean 8 to 20 second clip with one speaker and low background noise
• use Speaker Prep before blaming the model

There is also:

• a benchmark helper in backend/scripts/quality_benchmark.py
• a listening review format in docs/research/LISTENING_FEEDBACK_SYNTAX.md
• a scripting guide in docs/research/INDEXTTS2_SCRIPTING_PLAYBOOK.md

User Manual And Walkthrough Videos

If you want a guided tour of the app before using it, start here:

• Full user manual with screenshots: docs/manual/USER_MANUAL.md
• Speaker Prep video: docs/assets/manual/videos/speaker-prep-tab.webm
• Conversation Workflow video: docs/assets/manual/videos/conversation-workflow-tab.webm
• Conversation Results video: docs/assets/manual/videos/conversation-results-tab.webm
• Timeline Editor video: docs/assets/manual/videos/timeline-editor-tab.webm

Example Script

SpeakerOne: I think we should test three versions before we keep the final line.
SpeakerTwo: Good. If one sounds rushed, regenerate it and compare again.
SpeakerThree: After that, move the best takes into the timeline and export the scene.

Important Directories

• shared/audio/speakers/ - live speaker reference audio used for cloning
• shared/audio/source_clips/ - raw clips for preparation or batch processing
• shared/audio/speakers_backups/ - backups of original speaker files
• shared/models/checkpoints/ - IndexTTS model files
• shared/audio/outputs/ - exported outputs
• shared/audio/temp_conversation_segments/ - per-line conversation audio
• shared/audio/uploads/ - temporary imported files
• shared/data/project_saves/ - saved conversation projects
• shared/data/timeline_projects/ - saved timeline projects
• frontend/ - browser UI
• backend/ - FastAPI app plus wrapped IndexTTS runtime
• docs/ - manuals, research notes, release docs, and supporting references
• tools/ - maintenance helpers, manual capture scripts, and debug utilities
• examples/ - reusable sample inputs and saved examples

Container CLI Usage

If you want a CLI-style run without teaching users a host Python setup, use the backend container:

docker compose -f docker/docker-compose.yml exec backend python backend/indextts/cli.py "Your text here" -v /app/shared/audio/speakers/YourVoice.wav -o output.wav --model_dir /app/shared/models/checkpoints -c /app/shared/models/checkpoints/config.yaml

Docs

• Docs index: docs/README.md
• User manual: docs/manual/USER_MANUAL.md
• Docker guide: docker/README.md
• Deployment guide: docs/deployment/DEPLOYMENT_GUIDE.md
• API summary: docs/api/API_README.md
• Known limitations: docs/project/KNOWN_LIMITATIONS.md
• Release readiness: docs/project/RELEASE_READINESS_STATUS.md
• Audio folder guide: shared/audio/README.md

Ready To Try It

If you want the shortest path from install to first result:

1. Start the stack with docker\start.bat
2. Open http://localhost:3000
3. Prep or add a voice
4. Generate a short conversation
5. Review, regenerate, and export

Credit

The underlying model technology, papers, and official pretrained checkpoints belong to the IndexTTS team. This repository packages those models into a more workflow-focused local application.

Acknowledgements

• index-tts/index-tts
• tortoise-tts
• XTTSv2
• BigVGAN
• maskgct