Farstax
← All Stacks

Application Stack · voice-ai

VoiceStudio

Run a private voice-production service on workspace compute and let people or agents create, transcribe and transform speech without metered hosted voice APIs for the local workflow.

Workflows the agent can manage

  • clone and design voices
  • generate text-to-speech audio
  • transcribe speech
  • dub and translate video
  • produce stories and audiobooks
  • operate voice workflows from MCP or an OpenAI-compatible API

Agent operability

Editorial agent-readiness estimate: 5/5. This reflects documented interfaces and proposed workflows, not a successful installation or a support promise. Assessment date not recorded; verify upstream documentation before use.

  • MCP — Built-in MCP server exposes speech generation, voice cloning, transcription, voice/language discovery and health tools. The MCP transport is not authenticated, so keep it private/trusted or behind an appropriate secure boundary rather than exposing /mcp to the public internet.
  • REST — Local REST/SSE/WebSocket APIs include OpenAI-compatible audio endpoints. Remote HTTP/WebSocket access supports OMNIVOICE_API_KEY; use Bearer authentication over TLS/Tailscale/private reverse-proxy transport.

Evidence consulted

  • Official README documents Docker support; local REST/SSE/WebSocket and OpenAI-compatible audio APIs; MCP; local-first storage; active-beta status; and the stated RAM, disk, GPU and VRAM requirements. [stated; https://github.com/debpalash/VoiceStudio/blob/main/README.md]
  • Official Docker guide documents headless browser/server deployment, linux/amd64-only images, stable/exact release tags, persistent data/model-cache volumes, and OMNIVOICE_API_KEY administration. [stated; https://github.com/debpalash/VoiceStudio/blob/main/docs/install/docker.md]
  • Official MCP guide documents generate_speech, clone_voice, transcribe, discovery and health tools and states that the HTTP MCP transport is unauthenticated and should remain on a trusted LAN or behind TLS. [stated; https://github.com/debpalash/VoiceStudio/blob/main/docs/mcp.md]
  • Official API-auth guide documents Bearer-key authentication for remote HTTP/WebSocket access and recommends encrypted transport for remote API credentials. [stated; https://github.com/debpalash/VoiceStudio/blob/main/docs/api-auth.md]

Implementation notes

Pin a stable or exact VoiceStudio Docker release on an adequately sized linux/amd64 workspace; persist application data and model cache; generate OMNIVOICE_API_KEY through Workspace Secrets; keep the UI/API behind workspace TLS/private routing; keep /mcp internal rather than public; configure MCP file mode/base-path confinement when agents handle audio by path; and verify the licence of every selected model before enabling commercial workflows.

Licence / edition: AGPL-3.0-only application; model weights and third-party assets retain separate upstream licences Upstream labels VoiceStudio an active beta; pin a stable or exact version instead of the rolling latest/main Docker channel for a Farstax deployment. The application licence explicitly allows commercial and internal business use, but downloaded models are independently licensed and can impose non-commercial or other restrictions. The default k2-fsa/OmniVoice pretrained weights are identified upstream as CC-BY-NC; select and verify a commercially compatible model before commercial production use. Published Docker images are linux/amd64 only at qualification time.

Resources: heavy class — Upstream lists 8 GB RAM and 10 GB free disk as the default local-workflow minimum, with 16 GB+ RAM and 20 GB+ SSD recommended. GPU acceleration is optional; upstream lists 4 GB VRAM when using a GPU and recommends 8 GB+. CPU mode is supported but heavy generations can be slow. Published Docker images are currently linux/amd64 only, and first run downloads roughly 2.4 GB of default model weights. (stated)

External prerequisites: workspace with at least the upstream minimum resources; 16 GB+ RAM is the recommended target; linux/amd64 when using the published Docker image; sufficient persistent disk for application data, model weights and generated audio/video; explicit review of the chosen model licences before commercial use; optional compatible GPU/device runtime when acceleration is required

Security: Generate a strong OMNIVOICE_API_KEY for Docker/server mode and keep the master credential in workspace secret storage rather than chat or URLs. Use TLS, Tailscale or an equivalent private encrypted boundary for non-loopback API access; upstream warns that plain HTTP exposes Bearer credentials on hostile networks. Do not expose the built-in /mcp transport directly to the open internet: upstream documents it as unauthenticated and recommends trusted-LAN or TLS/private placement. Voice cloning can impersonate real people. Use consent-verified reference audio and apply the user's legal/policy obligations for generated speech and media. Catalogue qualification is not a security certification; inspect current advisories, supported versions, container guidance and upstream model supply-chain terms before setup.