Local • Free • Offline
Professional-grade voice processing in a single tool.

Jump to: Features · What Can VODER Do? · Quick Start · Modes at a Glance · Models Behind VODER · System Requirements · Documentation · Contributing

VODER brings together 8 processing modes under one interface — speech-to-text, text-to-speech, voice conversion, music generation, speech enhancement, sound effects, vocal separation, and speaker diarization — plus language dubbing (tts dub), any-to-any translation via TranslateGemma 12B, transcribe-edit-resynthesize (built into TTS interactive), three task-layer features that build on top of the modes: train (save reusable voice clones as .tts / .ttse), side-quests (quest — lightweight utility tasks like URL download and audio manipulation), and chains (user-defined pipelines that wire any number of voder oneline tasks together end-to-end) — and a natural-language AI agent, VADAR, that takes plain-English requests and runs any combination of the above for you. It runs entirely on your machine, needs no subscription, and works with or without a GPU.

• Multi-Speaker Dialogue System — Write scripts with multiple characters, each with a distinct voice. Control per-line timing, volume, and duration with script directives. Embed sound effects directly into dialogue lines and generate automatic background music that matches the spoken duration.
• Voice Design & Cloning — Describe a voice in plain English and VODER generates it, or provide a reference clip to clone a speaker's voice. Mix designed and cloned voices within the same dialogue.
• Speaker Separation — Extract individual speakers from multi-speaker recordings into separate audio files, each with a speaker-labeled transcript.
• Voice Conversion with Video I/O — Transform one voice into another while preserving words, emotion, and timing. Drop in an MP4 and get back a video with the converted voice.
• Music Generation & Manipulation — Generate full songs from lyrics and style descriptions. Remix, repaint, complete, extract stems, build individual instrument tracks, or replace background music in existing audio/video. Output up to 12 separate instrument tracks.
• Speech-to-Text with Intelligence — Transcribe audio, video, images, or direct URLs. Translate to any of 76 languages via TranslateGemma. Identify who spoke when with speaker diarization. Batch process multiple files.
• Language Dubbing — Translate speech from one language to another while preserving the original speaker's voice identity. Dub entire videos with per-segment timing alignment and background music preservation.
• Any-to-Any Translation — Translate between any of 76 languages using TranslateGemma 12B via the translate (source-target) syntax, decoupled from the ASR engine.
• Voice Re-Synthesis — Transcribe speech and re-read it in a different voice using tts svc, with an optional sts: prefix for high-fidelity voice conversion via Seed-VC v2.
• Side-Quests — Lightweight utility tasks that live outside the main engine: URL download, audio format conversion, cutting / merging / mixing / removing ranges, silence stripping, speed / pitch / soundlevel / bassboost / reverb / loudnorm effects, and more. Run python voder.py quest to see all available quests, grouped by category.
• Chains — User-defined pipelines that wire any number of voder tasks together: each chain is named, its output is captured to temp, and later chains can reference earlier chain names as input paths. Build a song, isolate its vocals, train a voice from them, then dub a video — all in one command.
• VADAR AI Agent — A natural-language agent powered by Gemma 4 12B (abliterated uncensored). Describe a task in plain English and VADAR thinks, decides, replies, and acts — running the right VODER commands in the right order, reading their outputs, and reporting results. Has its own tools (look, listen, watch, read, list, search, memory_*, calculate), session logging, persistent memories, and a configurable personality. No network access, no system shell — only VODER commands and project files.
• Smart Input Pipeline — Paste a YouTube, TikTok, Bilibili, Snapchat, Instagram, Facebook, or X/Twitter URL directly as input. VODER verifies the link actually points to a video before downloading. Feed an image and VODER extracts text via OCR. Automatically extract voice clips from multi-speaker audio for one-click voice cloning.

Describe a voice in plain English — "deep male voice, authoritative" — and VODER generates speech that matches. Or provide a reference audio clip and VODER clones the speaker's voice from it. Both approaches can be mixed in the same dialogue: some characters designed, others cloned.

Write scripts with multiple characters, each with a distinct voice. VODER assembles the full dialogue into a single audio file with per-line control over timing, volume, and duration via script directives (/time, /level, /duration). Embed sound effects directly into dialogue lines using the special sfx: character — door creaks, applause, rain — generated on the fly from text descriptions.

When generating dialogue, VODER can produce a background music track that exactly matches the spoken duration, mixed at a configurable volume with fade transitions. An optional reference (audio, video, or URL) can be provided for stylistic guidance — the reference is processed through SVS to extract clean instrumental before use. No manual editing or external tools needed.

Transform one voice into another while preserving the original words, emotion, and timing. Supports video input/output — drop in an MP4 and get back a video with the converted voice. For music, VODER switches to a high-fidelity 44.1kHz model. A mimic mode transfers not just the voice timbre but the accent and speaking style as well.

Generate full songs from lyrics and style descriptions. Beyond basic generation, VODER supports 6 sub-tasks: remix (style transfer with bias control), repaint (restyle a specific time range), complete (add missing instruments), lego (build individual tracks), extract (isolate specific stems), and bgm (replace background music in existing audio/video with generated music at a configurable volume). Output up to 12 individual instrument tracks for post-production. A three-tier quality system lets you trade speed for output quality.

Isolate clean vocals from any song, or extract the instrumental. Works with audio files, videos, and direct URLs from any supported platform (YouTube, TikTok, Bilibili, Snapchat, Instagram, Facebook, X/Twitter). This separation engine also runs automatically behind the scenes in TTS (to clean voice cloning references), STS (to improve conversion quality), and STT (to pre-clean audio before transcription).

Transcribe audio, video, images, or direct URLs to text. Supports translation to English from 99 languages, speaker diarization (who spoke when), and batch processing of multiple files. An overdose mode using Microsoft VibeVoice ASR delivers higher-quality transcription with built-in speaker identification.

Remove noise, reduce room echo, and restore clarity from degraded recordings. Upscale audio to 48kHz with AudioSR super-resolution (basic model for general audio, speech model for voice). Works on audio and video files alike.

Extract individual speakers from multi-speaker recordings into separate audio files, each with a speaker-labeled transcript. The blend flag preserves non-vocals (background audio) in each speaker's output, and the video flag muxes separated audio with the original video for MP4 output — useful for removing unwanted speakers from a video while keeping the visuals.

Translate speech from any language to English while preserving the original speaker's voice identity — tts slc "audio.wav". Supports any-to-any translation via TranslateGemma with the translate (source-target) syntax. An optional music flag preserves the original instrumental track, and overdose adds a voice fidelity pass. Accepts audio files, videos, and supported platform URLs.

Dub entire videos to another language with per-segment timing alignment — tts dub "video.mp4" auto-translates to English by default. Uses VibeVoice ASR with audio events, TranslateGemma per-segment translation, Fish S2 Pro voice cloning, speed adjustment, and timeline assembly. Add subtitle to burn translated subtitles. Add translate "(auto-ja)" to target any language. Preserves background music.

Paste a URL from YouTube, TikTok, Bilibili, Snapchat, Instagram, Facebook, or X/Twitter directly as input — VODER auto-detects the platform, verifies the link points to a real video (not a channel page, profile, photo post, or playlist), then downloads and processes it. Feed an image containing text and VODER extracts it via OCR for TTS processing. Automatic voice clip extraction from multi-speaker audio enables one-click voice cloning for dialogue characters.

Lightweight utility tasks that live outside the voder engine but are still useful in a voice-processing workflow — URL download, audio format conversion, cutting, merging, mixing (overlay multiple sources at specified start times), range removal, silence stripping, soundlevel / speed / pitch / bassboost / reverb / loudnorm, and more. Side-quests are grouped by category in the listing — download stands alone, the rest live under Media Manipulation. Run python voder.py quest (no args) to list every available side-quest with its one-line description. See COMMAND_CATALOG.md for the full list.

User-defined pipelines that wire any number of voder oneline tasks together. Each chain is named, runs a full voder oneline command, and its output is captured to a temp directory. Later chains can reference earlier chain names as input paths — VODER resolves them internally to the captured temp file. Side-quests work inside chains just like any other oneline task.

# TTM → SVS → STS pipeline
python src/voder.py chains "song" ttm lyrics "la la la" styling "pop" 30 / "voice" svs voice "song" / "cover" sts base "voice" target "ref.wav"

Use / (space slash space) to separate chains. Intermediate chain outputs live in temp_chains/; only the last non-empty chain's output reaches results/. Empty chains are skipped (their names remain available for reuse); duplicate names cause an error and stop the pipeline. This lets you compose pipelines that VODER's built-in modes never anticipated — generate music, isolate vocals, train a voice from them, then dub a video, all in a single command.

VADAR is the natural-language layer on top of everything else. Instead of remembering oneline syntax, you describe the task in plain English and VADAR figures out which VODER commands to run, in what order, and reads their outputs to verify the result. It runs locally with no network access — only VODER project files and paths you provide are reachable.

# Oneline — describe the task, VADAR runs the right commands
python src/voder.py vadar "Generate a 30-second upbeat pop song about rain, then isolate its vocals"

# Interactive CLI — choose option 10
python src/voder.py cli

VADAR is powered by Gemma 4 12B (abliterated uncensored variant from OpenYourMind/gemma-4-12B-it-abliterated-uncensored). The model downloads automatically on first run (downloads ~24GB into src/models/checkpoints/vadar/ via huggingface_hub.snapshot_download). See READ.md for setup. Without the model in place, vadar prints setup instructions and exits. It has its own tools (look, listen, watch, read, list, search, memory_read/write/edit/delete, calculate), persistent memories in src/voders/vadars/memories/, session logs in src/voders/vadars/sessions/, and a configurable personality in src/voders/vadars/about/. See Guide.md for the full VADAR user guide.

git clone https://github.com/HAKORADev/VODER.git && cd VODER
pip install -r requirements.txt && pip install --upgrade protobuf==5.29.6

# Help
python src/voder.py

# GUI
python src/voder.py gui

# CLI (interactive)
python src/voder.py cli

# One-liner examples
python src/voder.py tts script "Hello world" voice "female, cheerful"
python src/voder.py stt "audio.wav" timestamp dialogue
python src/voder.py sts base "input.wav" target "voice.wav"
python src/voder.py ttm lyrics "Walking down the street" styling "upbeat pop" 30
python src/voder.py svs "song.mp3" voice
python src/voder.py ss "meeting.wav"
python src/voder.py tts slc "foreign_speech.wav"
python src/voder.py tts dub "video.mp4"
python src/voder.py tts dub translate "(auto-ja)" "video.mp4"
python src/voder.py tts svc "speech.wav" target "voice_ref.wav"
python src/voder.py se "noisy_recording.wav"
python src/voder.py sfx sound "thunder rumbling" duration 10

# Side-quests (URL download & utility tasks — see `python voder.py quest`)
python src/voder.py quest download "https://youtube.com/watch?v=..."

# Voice training (save a voice clone for reuse in TTS)
python src/voder.py train voice:narrator "ref1.wav" "ref2.wav"
python src/voder.py train extreme voice:narrator "ref1.wav"

# Chains (wire multiple voder oneline tasks together)
python src/voder.py chains "song" ttm lyrics "la la la" styling "pop" 30 / "voice" svs voice "song" / "cover" sts base "voice" target "ref.wav"

# VADAR AI agent (describe a task in natural language, it decides what to run)
python src/voder.py vadar "Generate a 30-second upbeat pop song about rain, then isolate its vocals"

Run in Colab — no installation needed: Open in Google Colab

FFmpeg is required for audio processing. Install via your system package manager. See READ.md for all setup details.

VODER has 8 main processing modes — the engine's primary audio transformation pipelines. On top of these, three additional tasks & features layer utility workflows: voice training, side-quests, and chains. Sitting above all of them is VADAR, a natural-language AI agent that can call any of the modes or features on your behalf.

VODER orchestrates state-of-the-art open-source models — each selected for quality:

Some modes (SS, TTM overdose, ACE-Step complete) benefit from 24-32 GB VRAM or 48 GB+ system memory. See Guide.md for the full per-mode breakdown.

Speaker diarization requires a free Hugging Face token — set HF_TOKEN env var or HF_TOKEN.txt. See READ.md for details.

VODER is open-source under AGPL-3.0. Pull requests are not accepted — the codebase is maintained by a single developer. However, issues, bug reports, and feature suggestions are very welcome. If you want to build on top of VODER, fork it — that's what it's for.

Built for the open-source AI voice community.