Audio Clone Studio

Audio Clone Studio is a local-first voice cloning and speech production workspace built on top of the F5-TTS stack. It combines fast reference-based synthesis, project-aware asset management, trained voice checkpoints, reusable voice profiles, batch rendering, aligned speech editing, diagnostics, and a polished browser studio in one repository.

This repository is designed for practical production work:

• clone and render from curated reference clips
• manage reusable voice profiles and style prompts
• switch between shipped base models and local trained checkpoints
• queue preview and final renders safely on shared hardware
• perform transcript-aligned speech edits instead of destructive full rerenders
• review takes with diagnostics, comparison views, and export bundles
• serve the studio locally or mount it behind a small FastAPI service

Highlights

• Voice Studio: project-based browser workspace for references, styles, profiles, takes, and exports
• Trained Voice Workflow: route renders through a local finetuned checkpoint with profile-based reference selection
• Voice Profiles: group multiple clean references so the app can choose the best clip per render
• Editing Tools: alignment-first replace, insert, and delete actions with localized rerendering
• Batch and Queueing: render long scripts one job at a time on constrained machines
• Diagnostics: transcript drift checks, silence checks, reference scoring, and take review helpers
• Apple Silicon Support: Apple-friendly transcription extras and an optional MLX backend for shipped base-model inference
• API Mode: optional FastAPI server with mounted studio UI and local REST endpoints

What’s Included

User-Facing Apps

• f5-tts_voice-studio The main browser studio for cloning, rendering, editing, and export.

• f5-tts_infer-gradio The broader inference demo with multi-style and multi-speaker utilities.

• f5-tts_studio-server FastAPI server that exposes the studio UI under /app and API routes under /api/v1.

Production Workflows

• short-form and long-form voice cloning
• trained checkpoint auditioning
• project-scoped pronunciation dictionaries
• style prompt reuse
• batch render pipelines
• take comparison and export packaging
• transcript-aligned speech editing

Installation

1. Create an environment

python -m venv .venv
source .venv/bin/activate

If you prefer Conda:

conda create -n audio-clone python=3.11
conda activate audio-clone

2. Install FFmpeg

FFmpeg is required for reference preparation, audio decoding, and several export paths.

macOS:

brew install ffmpeg

Ubuntu / Debian:

sudo apt-get update
sudo apt-get install ffmpeg

3. Install PyTorch for your machine

Apple Silicon:

pip install torch torchaudio

NVIDIA:

pip install torch==2.8.0+cu128 torchaudio==2.8.0+cu128 --extra-index-url https://download.pytorch.org/whl/cu128

4. Install the project

Studio install:

pip install -e ".[studio]"

Apple-Silicon studio install:

pip install -e ".[studio,apple_audio]"

Training extras:

pip install -e ".[train]"

Development extras:

pip install -e ".[dev]"

Quick Start

Launch the main studio

f5-tts_voice-studio

Launch the API server

f5-tts_studio-server

Launch the broader inference demo

f5-tts_infer-gradio

Command-line generation

f5-tts_infer-cli \
  --model F5TTS_v1_Base \
  --ref_audio "ref_audio.wav" \
  --ref_text "Reference transcript." \
  --gen_text "Text to render."

Studio Workflow

1. Create a project

Start with a project in the studio and keep related references, styles, renders, and exports together.

2. Save clean references

Add short, clean clips with confirmed transcripts. The studio can analyze and score them for later reuse.

3. Build a voice profile

Group several strong references from the same speaker. The studio can then pick the most suitable clip automatically for each render.

4. Add style prompts

Save alternate delivery examples for pacing, tone, or mood so they can be reused across scripts.

5. Render

Use:

• Quick Preview for faster iteration
• Final Render for higher-quality delivery
• Voices for direct text-to-speech with a saved profile or single reference
• Trained Voice for checkpoint-based rendering

6. Review and export

Compare takes, run diagnostics, and export WAV, MP3, transcripts, or packaged bundles.

Apple Silicon Notes

On Apple Silicon, the studio supports:

• mlx-whisper for local transcription
• optional Apple MLX F5-TTS base-model inference from the Runtime panel
• automatic fallback to PyTorch when a local finetuned checkpoint is selected

Recommended install:

pip install -e ".[studio,apple_audio]"

Local Data Locations

The studio stores project data outside the repository.

macOS:

• support data: ~/Library/Application Support/F5-TTS-Studio
• cache: ~/Library/Caches/F5-TTS-Studio

This keeps references, profiles, takes, exports, and the local library out of the working tree.

Sharing

For temporary remote access, the repository includes helper scripts for local tunnel workflows:

./scripts/voice_studio_quick_tunnel.sh
./scripts/voice_studio_zrok.sh

The studio can also enforce optional access controls through environment variables:

export F5_TTS_STUDIO_USERNAME=demo
export F5_TTS_STUDIO_PASSWORD=change-me
export F5_TTS_STUDIO_TOKEN=optional-api-token
export F5_TTS_MAX_UPLOAD_MB=64

Repository Layout

src/f5_tts/
  api.py                 Public Python API
  infer/                 Inference entrypoints and demos
  studio/                Voice Studio app, runtime, storage, API, editing
  model/                 Model and trainer code
  train/                 Finetuning entrypoints
  runtime/               Deployment integrations

scripts/                 Local launch and sharing helpers
testsuite/               Test coverage for studio and service flows
ckpts/                   Optional local checkpoints

Development Commands

Syntax check:

python -m py_compile src/f5_tts/api.py

Run targeted tests:

python -m unittest -v testsuite.test_studio_service

Run the full test suite:

python -m unittest discover -s testsuite

Training

Finetuning entrypoints are available when training extras are installed:

f5-tts_finetune-cli
f5-tts_finetune-gradio

For a deeper walkthrough of inference usage, examples, and editing tools, see:

• Inference Guide
• Runtime Deployment Notes