Open Source
Open Source

VoiceStudio: the local-first open-source voice studio

The GitHub repo debpalash/VoiceStudio gained +5104 stars in a single week (week of 2026-09-07) to about 24.6k total, topping that week's momentum charts as a local-first voice project (AGPL-3.0, Python, active on 2026-09-11). Its positioning fits one line: an open-source, fully local ElevenLabs alternative - voice cloning, voice design, video dubbing, dictation, transcription, audiobook creation, covering about 646 languages, with the local workflow needing no account, API key, subscription, or usage meter. The underrated design is that it is not one voice model but an engine-orchestration layer integrating 16 TTS and 11 ASR engines, hot-swappable; it runs across macOS/Windows/Linux/Docker and ships an OpenAI-compatible local speech API plus an MCP server. This piece maps the capability surface, the local-first privacy/cost divide, and the division of labor with the same-week cloud real-time GPT-Live-1 (VoiceStudio leans to batch dubbing/transcription, not real-time conversation), then names five real constraints: AGPL-3.0 commercial caveats, beta stability, the ongoing Electron rewrite, uneven engine quality, and not every engine being local or free.

Published September 13, 202610 min read
<!-- voicestudio-resource | open-source | VoiceStudio: the local-first open-source voice studio -->

What it is and why the momentum is staggering

VoiceStudio is the poster child of 2026's local-first voice wave. On GitHub at debpalash/VoiceStudio, it holds 24,588 stars and gained +5104 stars in the week of September 7, 2026. Created April 9, 2026 and last pushed September 11, 2026, it carries 3030 forks and only 25 open issues.

Its positioning fits one sentence: an open-source, fully local alternative to ElevenLabs. It packs voice cloning, voice design, video dubbing, dictation, transcription, and audiobook creation into one desktop app, spans a 646-language speech synthesis catalogue, and for the local workflow requires no account, no API key, no subscription, and no usage meter. Previously named OmniVoice-Studio, it now uses the VoiceStudio brand.

Why the interest now? Not technical flash, but timing: frustration with usage-billed cloud voice APIs that push data off-machine, plus the 2026 "run it on your own hardware" localism. The same week OpenAI shipped the cloud real-time GPT-Live-1, VoiceStudio sits on the opposite side, keeping compute, data, and cost on your machine. The cloud path is in /en/posts/gpt-live-1-release-hotspot; this piece stays local.

Two boundaries up front. The project labels itself Active beta and recommends the latest release for stable work. Its README banner states an Electron rewrite is ongoing and asks users not to file desktop-app-related issues or pull requests. The license is AGPL-3.0, aggressive for commercial use, examined later.

Capability map: what 16 TTS and 11 ASR engines really mean

The most underrated design is that VoiceStudio is not a single voice model but an engine orchestration layer. It integrates 16 TTS engines and 11 ASR engines in a model catalogue, switchable there or with Ctrl/Cmd+E. The "engine registry" architecture weaves different models into a capability mesh rather than betting on one.

TTS side, the 16 engines diverge in language coverage, cloning, instruction support, and license:

EngineLanguagesCloningInstructLicense
VoiceStudio (default, k2-fsa/OmniVoice)600+YesYesAGPL-3.0 app / Apache-2.0 code + CC-BY-NC weights
CosyVoice 39 + 18 dialectsYesYesApache-2.0
GPT-SoVITS5YesNoMIT
VoxCPM230YesYesApache-2.0
MOSS-TTS-Nano20YesNoApache-2.0
KittenTTSEnglish onlyNoNoMIT
MLX-AudioModel-dependentVariesVariesVaries
Sherpa-ONNX20+NoNoApache-2.0
IndexTTS 2.5ZH/EN/JA/ES/ARYesNoBilibili model license
OmniVoice GGUF600+YesYesAGPL-3.0 app / derivative model terms
OmniVoice (subprocess)600+YesYesAGPL-3.0 app / Apache-2.0 code + CC-BY-NC weights
PocketTTSEN/FR/DE/PT/IT/ESYesNoCC-BY-4.0, gated
Supertonic 331NoNoOpenRAIL-M
MOSS-TTS-v1.531YesNoApache-2.0
dots.tts24YesNoApache-2.0
Confucius4-TTS14YesNoApache-2.0

Three takeaways. First, the default engine (k2-fsa/OmniVoice) covers 600+ languages and carries the 646-language claim. Second, not every engine clones; engines without cloning cannot preserve a reference speaker in dubbing or pinned-voice batch jobs, and VoiceStudio rejects those jobs rather than silently switching, responsible for production. Third, licenses range from permissive MIT and Apache-2.0 to restricted CC-BY-NC, OpenRAIL-M, and the Bilibili model license, so "open-source app" is not "free to commercialize". More later.

ASR side, 11 engines, WhisperX default at ~100 languages and strong at dubbing, subtitles, word-level timing:

EngineLanguagesBest fit
WhisperX (default)~100Dubbing, subtitles, word-level timing
Faster-Whisper~100General cross-platform transcription
Faster-Whisper (isolated)~100Crash-isolated batch transcription
MLX Whisper~100Apple Silicon
PyTorch Whisper~100CUDA, MPS, and CPU fallback
Parakeet TDTEnglish + 25 EUFast CPU/CUDA transcription
Parakeet TDT v3 (MLX)25 EUApple Silicon dictation and word timestamps
MoonshineEnglishLow-power, low-latency ONNX
FunASR50+VAD and inline diarization
sherpa-onnx (live dictation)Model-dependentStreaming CPU dictation
OpenAI-compatible (configured server)Server-dependentLocal gigastt/Qwen3-ASR or a remote endpoint

"646 languages" needs correct reading: it is the TTS language catalogue total, and actual coverage and quality depend on the engine. Breadth comes from the default engine; fine-grained quality varies by engine.

Cross-platform is another ace: macOS 13.3+ (Apple Silicon), Windows 10/11 x64, Linux x86_64 (glibc 2.39+), and Docker. Compute spans CUDA, Apple Silicon MPS/MLX, ROCm on Linux, CPU, and optional remote workers. One-line Docker:

bash
docker run -d -p 127.0.0.1:3900:3900 -v omnivoice-data:/app/omnivoice_data \
  --name voicestudio palashdeb/omnivoice-studio:stable

Architecture top-down: a Tauri v2 desktop shell (Rust) over IPC to a React + Vite UI; the UI talks to localhost:3900 over HTTP, SSE, WebSocket; behind it a FastAPI backend with TTS/ASR engine registries, dubbing/audio/long-form pipelines, OpenAI-compatible API and MCP server, and a SQLite + Alembic database (omnivoice_data/). Engine adapters isolate under backend/engines/, remote compute through a worker system, the foundation that absorbs 16 + 11 engines without contamination.

Local-first: the privacy and cost divide versus cloud APIs

The core judgment: the divide between VoiceStudio and the same-week cloud real-time API is not "whose voice is more natural" but where data and money flow.

Cost first. Cloud services (ElevenLabs or GPT-Live-1) share a pattern: register, get an API key, pay per call or burn credits. For short-video dubbing, batch audiobooks, internal transcription, the high-throughput but latency-tolerant work, the bill scales linearly with usage. VoiceStudio's app is free software; you supply the hardware. Cost model: buy a GPU once or reuse existing machines, then unlimited calls with no marginal fee. For production workloads, this ceiling is far friendlier.

Data and privacy. VoiceStudio's default path is local: voices, projects, settings, outputs stay on your machine, network features are explicit opt-ins. The README states that for the default local workflow, recordings, transcripts, voices, projects remain strictly on local disk, leaving only if you configure remote workers or external ASR endpoints. It also integrates AudioSeal imperceptible watermarking by default to detect synthetic speech without altering quality.

Contrast with cloud: cloud processes audio and text on provider servers; alongside per-usage billing, your data leaves the machine. We do not deny cloud's zero-setup and elastic scaling, but for internal meetings, client assets, unreleased recordings, local "data never leaves" is a hard constraint. Full trade-off in /en/posts/cloud-vs-local-voice-ai-review.

A missed detail is the network boundary: the desktop only talks to a loopback backend on localhost:3900, loopback calls need no server key, remote access needs a share PIN or API key. Loopback ASR may use HTTP and keeps audio local; non-loopback endpoints force HTTPS and follow no redirects. Analytics is off by default and even when on sends only allowlisted, content-free metadata, never text, audio, file names, or projects. This makes "local-first" more than a slogan.

Positioning: batch dubbing and transcription, not real-time interaction

In the larger map, VoiceStudio leans toward production-side batch processing and long-form content, not interaction-side real-time conversation.

Workflows include voice cloning and design, video dubbing, dictation, stories and audiobooks, batch generation, all queueable, waitable, producing deliverable files. Dubbing supports file or URL import, captions, optional YouTube sign-in, combined waveform and timed-transcript editing, multi-speaker preservation. Audiobooks support EPUB/PDF import, chapter rendering, .m4b export. The batch queue watches a local folder and auto-processes new videos, a clear "content factory" pipeline.

The same-week GPT-Live-1 cloud API targets low-latency, bidirectional, conversational interaction: phone support, live companionship, instant voice-agent replies. Overlap is small. VoiceStudio exposes a live transcription endpoint (WS /v1/audio/transcriptions/stream) but its battlefield is not "talking with users in real time" but "turning audio and text into finished products".

Conclusion: for a real-time phone voice agent, VoiceStudio is not built for it (see FAQ); for multilingual tracks on a hundred videos, a book into an audiobook, meeting audio into speaker-labeled text, it is among the handiest open-source choices. Local API invocation pattern in /en/posts/gpt-live-1-voice-api-sop (cloud piece, but API thinking transfers).

It also mounts an MCP server (http://localhost:3900/mcp) so Claude Desktop, Cursor, and agents call generate_speech, clone_voice, transcribe, plus npx skills add debpalash/VoiceStudio for compatible agents — local voice infrastructure, not just a desktop app.

A sober look: five real constraints

Heat aside, honesty about constraints.

First, AGPL-3.0 commercial caveats. The app lets you run, modify, use internally, and does not restrict selling generated audio; but "downloaded model and tokenizer terms may". Default OmniVoice weights are CC-BY-NC (non-commercial) with a tokenizer under separate terms. AGPL's network clause: modify and offer as a network service, you must provide corresponding source under the same license. A commercial license exists for VoiceStudio-owned code but does not relicense third-party models. Individuals and internal use are safe; external service or closed-source integration must review model licenses engine by engine.

Second, beta stability. Active beta, main may change between releases, pin to latest release. Production means accepting occasional bugs, reading issues, running self-checks (uv run python backend/main.py --diagnose --deep).

Third, Electron rewrite risk. The README banner says an Electron rewrite is ongoing and asks users not to file desktop-app issues or PRs. The current desktop shell (Tauri v2 / Rust) is being replaced; desktop experience may change substantially, related bugs unlikely prioritized soon. API and batch users are minimally affected; heavy desktop users should brace.

Fourth, uneven engine quality. Among 16 TTS engines, some target high-fidelity zero-shot cloning (OmniVoice, CosyVoice 3), others are lightweight English engines (KittenTTS); coverage ranges 5 to 600+, licenses from MIT to restricted. Wrong engine yields a rejected clone or mismatched timbre. VoiceStudio gives hardware-based pairings (Apple Silicon: MLX-Audio/OmniVoice + MLX Whisper; NVIDIA 8GB+: OmniVoice/CosyVoice 3 + WhisperX; low-VRAM/CPU: PocketTTS/Sherpa-ONNX/KittenTTS + Moonshine/Faster-Whisper int8), but tuning the right engine is itself a threshold.

Fifth, not every engine is local or free. The 646-language catalogue rests on local weights, but some engines (PocketTTS) need gated access, and IndexTTS 2.5 needs a separate written license above 100M monthly active users or RMB 1B annual revenue; the OpenAI-compatible ASR can point to a remote endpoint, audio then leaves local. The Google Colab notebook for install-free trials runs on remote compute, so audio does not stay local. "Fully local, fully free" is a state you must configure and defend, not a default.

Collapsing the five: VoiceStudio's value is real, especially for production-heavy, privacy-sensitive teams wanting to escape per-usage billing, but it is not "install and done". It is a local voice operating system demanding you understand licenses, manage engines, accept beta fluctuation.

FAQ

Q1: Can it run fully offline?

A1: In the default local workflow, yes. Voices, projects, settings, outputs stay on your machine, loopback API calls need no server key. Two premises: you must download required model weights online at least once (first launch creates a managed Python environment and downloads the default model); data leaves local only if you configure remote workers or external ASR endpoints. The install-free Google Colab notebook runs on remote compute, so audio does not remain local.

Q2: Can I use it commercially under AGPL-3.0?

A2: The application does not restrict selling generated audio and allows internal use and modification, but watch two things. Downloaded models and tokenizers carry their own licenses, for instance default OmniVoice weights are CC-BY-NC, so review model terms engine by engine before commercial use. AGPL's strong copyleft: modify and offer as a network service, you must provide corresponding source under the same license. A commercial license covers VoiceStudio-owned code for proprietary embedding but does not relicense third-party models. Personal and internal use are safe; external or closed-source integration needs caution.

Q3: How is it fundamentally different from ElevenLabs?

A3: The core difference is the flow of data and cost. ElevenLabs is a hosted cloud service requiring an account, an API key, and per-usage billing, with audio and text processed on provider servers. VoiceStudio is an open-source, fully local alternative whose local workflow needs no account, key, subscription, or usage meter, and data stays off-network by default. In capability, VoiceStudio spans 646 languages through 16 TTS and 11 ASR engines and leans toward batch dubbing, transcription, long-form production, while ElevenLabs is more convenient in zero-setup and managed elasticity. For easy cloud use pick ElevenLabs; for controllable low-cost local use pick VoiceStudio.

Q4: What hardware does it require?

A4: Minimum is Windows 10 x64 / macOS 13.3 Apple Silicon / Linux x86_64 (glibc 2.39+), 8 GB RAM, 10 GB free disk, GPU optional with CPU mode supported; with a GPU at least 4 GB VRAM, recommended 8 GB+, large optional engines may need 12 to 16 GB or more. Recommended is 16 GB+ RAM, 20 GB+ SSD, NVIDIA CUDA or Apple Silicon. Intel Macs cannot run the local backend (missing PyTorch wheels) and only connect to a remote backend. Low-VRAM or CPU-only machines should choose lightweight engines such as PocketTTS, Sherpa-ONNX, KittenTTS, Moonshine, or Faster-Whisper (int8).

Q5: Can it power real-time phone voice or a real-time voice agent?

A5: This is not its main battlefield. VoiceStudio leans toward production-side batch processing and long-form content: video dubbing, audiobooks, batch transcription, dictation, and its live transcription endpoint (WS /v1/audio/transcriptions/stream) mainly serves dictation. It is not designed for low-latency, bidirectional, conversational real-time phone interaction such as phone support or instant voice-agent responses; that fits the same-week cloud real-time speech API better (see /en/posts/gpt-live-1-release-hotspot). If you need to turn piles of audio into finished products, it is a great choice; if you need to talk with users in real time, pick another solution.

This article is AI-assisted and human-edited. Last updated: 2026-09-13

FAQ

Can it run fully offline?
In the default local workflow, yes. Voices, projects, settings, outputs stay on your machine, loopback API calls need no server key. Two premises: you must download required model weights online at least once (first launch creates a managed Python environment and downloads the default model); data leaves local only if you configure remote workers or external ASR endpoints. The install-free Google Colab notebook runs on remote compute, so audio does not remain local.
Can I use it commercially under AGPL-3.0?
The application does not restrict selling generated audio and allows internal use and modification, but watch two things. Downloaded models and tokenizers carry their own licenses, for instance default OmniVoice weights are CC-BY-NC, so review model terms engine by engine before commercial use. AGPL's strong copyleft: modify and offer as a network service, you must provide corresponding source under the same license. A commercial license covers VoiceStudio-owned code for proprietary embedding but does not relicense third-party models. Personal and internal use are safe; external or closed-source integration needs caution.
How is it fundamentally different from ElevenLabs?
The core difference is the flow of data and cost. ElevenLabs is a hosted cloud service requiring an account, an API key, and per-usage billing, with audio and text processed on provider servers. VoiceStudio is an open-source, fully local alternative whose local workflow needs no account, key, subscription, or usage meter, and data stays off-network by default. In capability, VoiceStudio spans 646 languages through 16 TTS and 11 ASR engines and leans toward batch dubbing, transcription, long-form production, while ElevenLabs is more convenient in zero-setup and managed elasticity. For easy cloud use pick ElevenLabs; for controllable low-cost local use pick VoiceStudio.
What hardware does it require?
Minimum is Windows 10 x64 / macOS 13.3 Apple Silicon / Linux x86_64 (glibc 2.39+), 8 GB RAM, 10 GB free disk, GPU optional with CPU mode supported; with a GPU at least 4 GB VRAM, recommended 8 GB+, large optional engines may need 12 to 16 GB or more. Recommended is 16 GB+ RAM, 20 GB+ SSD, NVIDIA CUDA or Apple Silicon. Intel Macs cannot run the local backend (missing PyTorch wheels) and only connect to a remote backend. Low-VRAM or CPU-only machines should choose lightweight engines such as PocketTTS, Sherpa-ONNX, KittenTTS, Moonshine, or Faster-Whisper (int8).
Can it power real-time phone voice or a real-time voice agent?
This is not its main battlefield. VoiceStudio leans toward production-side batch processing and long-form content: video dubbing, audiobooks, batch transcription, dictation, and its live transcription endpoint (`WS /v1/audio/transcriptions/stream`) mainly serves dictation. It is not designed for low-latency, bidirectional, conversational real-time phone interaction such as phone support or instant voice-agent responses; that fits the same-week cloud real-time speech API better (see `/en/posts/gpt-live-1-release-hotspot`). If you need to turn piles of audio into finished products, it is a great choice; if you need to talk with users in real time, pick another solution.

Related

Open Source

LLaDA-Image: Ant Full-Open 6B Unified Image Generation Model

Ant Group's InclusionAI open-sourced LLaDA-Image, a 6B unified image generation and editing model (208 stars / Python / created 2026-08-31, snapshot 2026-09-09). One checkpoint does both text-to-image and instruction-guided editing; both backbone and DiT are diffusion models trained in a unified framework, with image-only pre-training establishing the visual prior; the Turbo variant uses Twin-DMD distillation to cut 50 steps down to 4. It scores 53.53 (English) and 53.38 (Chinese) on Qwen-Image-Bench, a double SOTA. HuggingFace and ModelScope host Base and Turbo weights, each with an FP8 variant, and community ComfyUI support landed on 2026-09-07. Biggest caveat: the repo's license field is null with no LICENSE file - confirm terms with InclusionAI before commercial use rather than assuming Apache-2.0 or MIT.

Sep 9, 202610 min read
Open Source

DeepSeek Harness: A Plugin-Everything Agent Framework

DeepSeek open-sourced its agent orchestration framework DeepSeek Harness (CLI: dsh) on GitHub under MIT, written in TypeScript and built on the Cordis runtime with an "everything-is-a-plugin" architecture that modularly assembles AI pipelines. The repo was created 2026-08-13 and passed 200k stars within ~3 weeks; it is currently 0.1.3-alpha, a developer preview with breaking changes expected (read SAFETY.md first). Launch the Web UI with `npx @deepseek-ai/dsh web` at http://127.0.0.1:3080.

Sep 5, 202610 min read
Open Source

Architecture Innovation Needs Someone to Write the Kernels: Qwen Open-Sources FlashQLA, 2-3x Faster GDN Forward, Hopper and Up Only

QwenLM/FlashQLA (GitHub API, checked 2026-08-30: 670 stars / 69 forks, Python, MIT, created 2026-04-24, last push 2026-08-26) is a high-performance linear attention kernel library built on TileLang. The authors report 2-3x forward and 2x backward speedups for GDN Chunked Prefill over the FLA Triton kernel on NVIDIA Hopper and Blackwell, with the largest gains in pretraining and edge-side agentic inference. Three things worth studying: gate-driven automatic intra-card context parallelism (exploiting the GDN gate's exponential decay under TP, long sequences and small head counts), hardware-friendly algebraic reformulation (cutting Tensor Core, CUDA Core and SFU overhead without losing precision), and TileLang fused warp-specialized kernels (hand-written warpgroup specialization that balances context-parallel and backward needs). Since v0.1.2 it also serves as a GDN backend for flash-linear-attention, plug-and-play through the standard FLA API. The timing is the interesting part: it shipped the same day as Qwen3.8-Flash-Next, meaning Alibaba handed in the architecture paper and the kernel implementation together. The entry ticket is not cheap though - SM90 or newer, CUDA 12.8+, PyTorch 2.8+, and the 2-3x figure is self-reported rather than independently reproduced.

Aug 30, 20268 min read