Local-first workflow
Transcripts, reference voices, and generated audio stay on your machine.
Local voice generation, composed
Voice cloning.
Local by default.
A terminal-native CLI that turns long-form transcripts into speech with reusable voices, rich diagnostics, and a workflow that stays on your machine.
❯
Local-firstOpen sourceLong-form readyCUDA accelerationPrivacy focusedWindows CLI
Stable release v0.1.7
Ready for WindowsAn isolated runtime. One reliable installer.
LLMVoice selects a CUDA or CPU profile automatically, keeps global Python untouched, and adds only its own bin directory to your user PATH.
- 1Install FFmpeg Shared
- 2Download the LLMVoice installer
- 3Run the installer
- 4Open a new terminal
- 5Run llmvoice doctor
Installation guide SHA256-verified release
01 · FFmpeg Shared
winget install --id Gyan.FFmpeg.Shared -e02 · Download and run installer
irm https://llmvoice.yusufdere.com/install.ps1 -OutFile install.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps103 · Validate installation
llmvoice doctorThe installer is downloaded, then executed separately. A remote script is never piped directly into execution.
Purpose-built workflowEverything the generation loop needs.
Everything the generation loop needs.
Nothing it doesn’t.
LLMVoice keeps the surface small while making the local workflow legible, resilient, and pleasant to use.
Reference voice cloning
Add and reuse named voices from a clean reference WAV in one command.
Long-form ready
Chunking, pauses, merging, and encoding are handled as one reliable pipeline.
Clean CLI UX
Readable output, purposeful progress, and errors that tell you what to fix.
NVIDIA CUDA or CPU
Uses NVIDIA CUDA acceleration when available, with CPU fallback for supported systems.
Diagnostics built in
Doctor, dry-run, and debug modes make the local environment inspectable.
Open architecture
MIT-licensed source with a focused Python CLI surface and no service layer.
Safe output handling
Explicit overwrite behavior, disk checks, and predictable output paths.
Configurable by design
Set voice, language, device, speed, chunk size, and pause defaults.
Four deliberate steps
Reference in. Finished audio out.
01
Add a voice
Create a named local voice from clean reference audio.
02
Feed a transcript
Point the CLI at a text file, short or long.
03
Generate locally
XTTS-v2 runs through CUDA or a CPU fallback.
04
Get your MP3
Chunks merge into one clean, ready-to-play file.
Terminal-native
A workflow you can read at a glance.
Inspect the machine, register a voice, set a default, then generate. Each command does one thing and reports exactly what happened.
Explore every command workflow.sh ready
01
$ llmvoice doctorEnvironment ready02
$ llmvoice voice add friday reference.wavVoice registered03
$ llmvoice config set default_voice fridayDefault updated04
$ llmvoice start transcript.txtGeneration startedPrivacy is the architecture
Your voice never needs to leave the room.
LLMVoice is a local CLI, not a hosted service. There is no account, ingestion pipeline, or remote inference layer between your files and your output.
The installer and first XTTS-v2 model download require network access. Voice references, transcripts, and generated audio are not uploaded for inference.
No cloud APIs
No telemetry
No transcript uploads
No voice uploads
Everything stays local
Documentation
Browse all docs From zero to a repeatable local workflow.
Know the license boundary
Open source code. Separately licensed model.
LLMVoice source code is MIT licensed. XTTS-v2 model weights are licensed separately under the Coqui Public Model License (CPML); the MIT license does not grant commercial rights to model usage. LLMVoice is positioned for local, non-commercial use unless you hold the appropriate model rights.
Read the licensing guideBuilt in the open.
Designed to stay local.
Inspect the source, understand the pipeline, and make the workflow your own.