VoiceStudio

Name: VoiceStudio — Download free. Local voice platform
Availability: InStock
Rating: 5 (1 reviews)
Author: Palash Debnath

VoiceStudio — Free Download. Local voice platform

VoiceStudio is a desktop application for voice cloning, video dubbing, dictation, and long-form audio generation that runs entirely on local hardware. It operates without an account, API key, subscription, or usage meter for the local workflow. The program supports zero-shot voice synthesis from short reference clips, voice design from textual descriptions, multi-speaker audiobook rendering, system-wide dictation, vocal isolation, speaker diarization, and batch processing of audio and video jobs. VoiceStudio includes a model catalogue for managing TTS, ASR, and LLM engines, remote model downloads to enrolled workers, GPU auto-detection across CUDA, MPS, ROCm, and CPU, an AI watermark system using AudioSeal, an MCP server for synthesis and transcription tools, diagnostics with self-checks and logs, and a local-first architecture where network-backed features are explicit opt-ins.

★★★★★

5.0(1 ratings)

Download VoiceStudio (Official links)

File size: 170 MB

The latest version of VoiceStudio is: 0.5.2

Operating system: Windows, Mac OS, Linux

Languages: English

Developer: Palash Debnath

Price: $0.00 USD (Open Source (AGPL-3.0))

Voice Cloning. Zero-shot synthesis from a short reference clip, typically three seconds is sufficient. The program mirrors the acoustic characteristics of the provided voice, including timbre and prosodic patterns, without requiring prior training on the specific speaker. The cloned reference persists to disk across restarts, so the first generation of a session skips re-encoding and any transcription pass. Ad-hoc references stay alive until cancelled or timed-out GPU work stops reading them, allowing prompt caching to finish without failing on a deleted temporary file. Voice cloning works with the default OmniVoice engine and supports combining style attributes with a reference clip, where the reference wins conflicts.
Voice Design. Build a new voice from a sentence by specifying gender, age, accent, pitch, emotion, and delivery instructions. The voice design feature creates a synthetic voice without a reference audio sample, using attribute-based synthesis to generate a consistent vocal identity. The default engine supports voice design from attributes alone, and consistent instruct stabilization improves cloning when style attributes are combined with a reference clip. Inline pronunciation control is available through pinyin or CMU phonemes, allowing precise phonetic adjustments for specific words or names.
Video Dubbing. Transcribe, translate, re-voice, and export video while preserving each speaker and aligning timing. The dubbing workflow separates spoken and target languages, preserves translations through segment cleanup, and allows failed translations to be retried or skipped without restarting the batch. Dubbing playback keeps an audible companion source when a WebView can render the preview picture but cannot decode its audio. Uploaded dubbing videos are normalized to browser-safe H.264/AAC before preview, preventing valid VP9, AV1, or Opus media from failing with no supported sources. Default-engine dubbing synthesizes several segments per forward pass instead of one call per line, with batch width following the host device headroom.
Stories and Audiobooks. Multi-voice scripts with EPUB and PDF import, chapter rendering, and M4B export. The feature casts characters to different cloned or designed voices and renders long-form content as a chaptered audiobook. Voices stay warm across restarts through persistent encoded references, so the first generation of a session skips re-encoding. Chapter rendering processes each chapter with its assigned voice set, and the M4B export produces a standard audiobook container compatible with major players.
Dictation Widget. System-wide shortcut activation, live transcription, and optional local LLM cleanup. The widget stays bound to the app where it started and recovers locally from silent recognizer output. A bundled Rust loopback sidecar exposes dictation start, stop, toggle, focused-output sessions, discovery, and JSON-RPC, while the backend adds versioned streaming events and a dependency-free CLI bridge for Herdr, coding agents, editors, desktop apps, and TUIs. Dictation carries one native output session from shortcut-down through final delivery and restores text, HTML, image, or file-list clipboards only when untouched.
Vocal Isolation. Demucs speech and background separation for extracting clean vocal tracks from mixed audio. The feature removes music, noise, and other background elements from recordings, leaving only the speech component. Isolation runs locally through the media processing engines and produces separate stems that can be used for further processing or export. Missing media-process failures point to repair controls, and first-run media-engine downloads become available to TTS immediately without a restart.
Speaker Diarization. Pyannote and WhisperX speaker assignment for identifying who spoke when in multi-speaker audio. The feature labels each speech segment with a speaker identity, enabling per-speaker processing in dubbing and transcription workflows. Diarization output feeds directly into the video dubbing pipeline, where each speaker is assigned a clone or designed voice for re-voicing. The speaker assignment remains stable across segment cleanup and translation preservation.
Batch Queue. Queue large sets of audio and video jobs with per-job progress, or watch a local folder for new videos. The queue processes jobs sequentially or in parallel depending on available device capacity, with per-engine batch width configurable from one to eight segments per forward pass. A timed-out in-process dub transcription no longer starts a second WhisperX or CTranslate2 call over the abandoned native worker, preventing overlapping access that preceded Windows 0xC0000005 exits. Importing replacement SRT subtitles keeps each cue bound to the best-overlapping source speaker and clone instead of resetting every line to a random default voice.
Model Catalogue. Install, remove, select, and route TTS, ASR, and LLM models from a unified interface. The catalogue uses one breathable workspace canvas with simpler pane and engine-family navigation instead of nested cards and scroll regions. Each engine has its own guide covering installation, configuration, and performance characteristics. Model-load GPU exhaustion returns a sanitized, actionable dubbing error, and readiness correctly attributes the shared model status to TTS. The catalogue rows use available desktop width and keep identity, runtime state, and actions from crowding one another.
Remote Model Downloads. Install models on enrolled remote workers with live progress tracking. The feature enables headless NVIDIA and ROCm machines to join as worker-only Docker Compose services with no published UI and durable protocol-v2 enrollment. The backend reports startup progress live, with a startup progress endpoint listing every step while PyTorch, API routes, and database migrations load in the background. Worker machines must update both the controller and worker together before reconnecting.
GPU Auto-Detect. CUDA, MPS, ROCm, and CPU routing with per-engine checks for optimal inference device selection. Settings include a compute-device override that pins the device when auto-detect picks wrong, offering only devices the machine actually has. Optional FlashInfer acceleration for the default engine on CUDA provides approximately 2.2x measured speedup when the optional flashinfer-python package is installed, with missing package or kernel failure logging why and falling back to the standard path. WSL2 AMD containers use the /dev/dxg ROCDXG bridge with actionable GPU diagnostics instead of silently falling back to CPU.
AI Watermark. AudioSeal embedding and detection for invisible watermarking of generated audio. The locally cached AudioSeal watermark generator warms on a background thread approximately 35 seconds after boot, so the first synthesis no longer serializes the audioseal import and model load inline. Invisible watermarking no longer stalls or silently skips the first take of a session. Watermark detection verifies the presence of the embedded signal, providing content provenance for generated speech.
MCP Server. Synthesis and transcription tools for MCP clients through versioned HTTP, WebSocket, JSON-RPC, CLI, and MCP transports. The MCP server enables other applications to trigger native dictation or connect programmatically for speech synthesis and transcription tasks. A dependency-free CLI bridge supports Herdr, coding agents, editors, desktop apps, and TUIs. The backend binds its port immediately and answers within a second of launch, narrating startup step by step.
Diagnostics. Self-checks, error journal, logs, and scrubbed support bundles for troubleshooting. The bug reporter notices when running an outdated build and offers the latest release before filing, with a file-anyway escape hatch and a Build status line stamped into every report. Windows debugger termination code 0x40010004 is no longer misreported as a backend crash or charged against automatic restart recovery. Repeated crash loops stop loudly with diagnostics, while source-mode development restarts an isolated backend crash without tearing down the UI.

VoiceStudio development began as a local-first alternative to cloud-based voice synthesis platforms, with the core architecture built in Rust for the backend and loopback sidecar, while the desktop shell and UI layer use web technologies rendered through a native container. The program supports macOS, Windows, and Linux, including Linux ARM64 for Asahi Apple Silicon with the OmniVoice GGUF engine built with GGML Vulkan where the toolchain allows. One-command installation is available on every desktop OS through a platform-detecting script. The Docker quick start requires an administrator key needed through container NAT instead of starting a UI whose protected actions return 403. Source installs on AMD GPUs honour the OMNIVOICE_TORCH_VARIANT environment variable to swap in the ROCm torch wheel after dependency synchronization. CI enforces performance regression budgets on hot paths, with operation-count tests pinning streaming TTS to one synthesis per sentence and cached dub re-mixes to zero re-synthesis. Linux source launchers catch missing libxdo and GStreamer audio plugins before they can cause a linker error or an aborted, blank WebKit renderer. The Model Catalogue and engine guides cover all 16 TTS and 11 ASR engines supported by the platform.