Requirements

Home · Manual index

Identity and requirements

Portable TTS Server V2 is a linux distro for windows users. Windows provides a native desktop window and CLI; inference and audio processing run inside this app's dedicated WSL2 distro. The complete release includes its own dependencies and Kokoro model data. Optional engines install inside its portable disk when chosen.

Its source project is aivrar/portable-linux-in-a-box. This TTS app originated as a child distro of Portable Linux in a Box. The distributed V2 Linux image is rebuilt from clean Ubuntu with the speech dependencies included.

Product identity and paths

Item	Location or behavior
Windows app root	Any writable folder on a local drive; E:\tts_server is an example
WSL distro	TTS-Server-V2-<12-character path hash>
Portable Linux disk	<app>\wsl\ext4.vhdx
Linux-private state	/opt/tts_server in that disk
Clean initial runtime	<app>\runtime\linux-rootfs.tar.gz
Start application	<app>\TTSServer.exe
Desktop host	<app>\runtime\launcher\TTSServer.exe
Browser runtime	<app>\runtime\webview2
Browser profile	<app>\cache\webview2
Windows CLI Python	<app>\runtime\python\python.exe
Start / stop	TTSServer.exe (or Start-TTSServer.cmd) / Stop-TTSServer.cmd
Discovery	<app>\output\run\registry\tts_server.json
Default ports	Bridge 9300, gateway 8300; customizable at startup

The legacy distro name linbox-TTS_Server is accepted only when its registered disk already belongs to this exact folder. Startup validates the WSL registration against the folder, rather than selecting a default Ubuntu distro. Linux-private paths stay within /opt/tts_server; host-visible paths stay within the app tree.

Host requirements

Windows and WSL2

Use Windows x64 with WSL2 and hardware virtualization enabled. Check wsl --status. If WSL is missing, install it from an administrator terminal with wsl --install --no-distribution, then restart if requested. WSL is a Windows component and cannot be made folder-local by this app. WSL1 and Windows ARM packages are not provided.

App dependencies are bundled

The release supplies Windows Python for tts.cmd, a self-contained .NET native launcher, WebView2 Fixed Version, and Linux Python, PyTorch, FFmpeg, espeak-ng, SoX, rubberband, and speech dependencies. No separate Python, .NET, WebView2, or CUDA toolkit install is required for the complete release. Source checkouts need the build procedure in release/BUILD.md.

Offline Kokoro and optional engines

Kokoro weights, voices, the English spaCy model, Japanese UniDic data, and Chinese pronunciation dependencies are included. Nine Kokoro language groups are checked by actual CPU synthesis with Python network connections blocked. No Hugging Face token is required for bundled Kokoro.

Choose other engines in Setup; their packages and weights download into this app's VHD. Gated engines require your own Hugging Face access. Whisper transcription/subtitle weights are optional downloads. Edge uses an online service for every request. Offline availability is specific to installed engines and their selected language/voice assets.

Folder and GPU

Extract the whole release into a writable local folder. Spaces are supported; UNC/network shares are not. The first start imports the bundled image into wsl/ext4.vhdx. Later starts use that disk. Run Stop before moving or copying it; see start and portability.

Kokoro can run on CPU. For GPU use, install a compatible Windows NVIDIA driver with WSL support. The CUDA user-space libraries are bundled, but a host GPU driver is not. GPU numbers vary by computer; choose a currently listed device.

Disk space for engines and weights

Weights and override packages live on the VHDX under /opt/tts_server/data and /opt/tts_server/overrides, not on the small Windows-visible folders. Plan space on the VHDX / the local volume that holds it. Published weights_size values from MODEL_SETUP (these are the installer-facing sizes, not VRAM):

Engine id	Display	Approximate weight size
edge	Edge TTS	none (cloud)
whisper	Whisper	~145MB base; other sizes optional
vits	VITS	~150MB
speecht5	SpeechT5	~250MB
kokoro	Kokoro 82M	~300MB
f5	F5-TTS	~1.35GB selected checkpoint
outetts	OuteTTS 1.0 0.6B	~1.22GB model + pinned 24kHz DAC
xtts	XTTS v2	~1.8GB
fish	Fish Speech	~3.4GB model + codec
parler	Parler-TTS	~3.6GB + tokenizer
voxcpm2	VoxCPM2	~4.96GB selected checkpoint
vibevoice	VibeVoice	~5.41GB checkpoint + local Qwen tokenizer
dia	Dia 1.6B	~6.45GB selected checkpoint + local DAC decoder
orpheus	Orpheus 3B	~6.72GB selected FP16 checkpoint + SNAC decoder
csm	Sesame CSM-1B	~7.15GB selected Transformers checkpoint
voxtral	Voxtral 4B TTS	~8.03GB selected checkpoint and preset embeddings
bark	Bark	~12GB + voice prompts
higgs	Higgs Audio 3B	~12GB model + audio tokenizer + HuBERT
qwen	Qwen Omni 7B	~21GB selected runtime snapshot
chatterbox	Chatterbox	English runtime assets only (pinned files)

Installing all local engines needs tens of gigabytes on the VHDX plus Hugging Face / pip caches under /opt/tts_server/cache. User audio (jobs, projects, voices) additionally grows E:\tts_server\output and E:\tts_server\projects_output. Use tts.cmd disk or GET /api/disk to see per-directory size and mount free space.

8. GPU vs CPU vs cloud, as the engines actually apply

VRAM planning estimates (MODEL_VRAM_ESTIMATE_GB) are conservative numbers for orchestrators, not hard CUDA limits. The gateway also keeps a safety reserve (TTS_SERVER_GPU_RESERVE_GB, default 1.5 GiB) and serializes loads per GPU.

Class	Engine ids	What you need
Cloud, no GPU, no weights	edge	Working internet. Live catalog of 300+ Microsoft neural voices.
CPU-friendly local	vits, speecht5, kokoro (Kokoro also runs well on GPU)	Modest RAM. VITS is documented as no GPU needed. Higgs lists CPU support in its Setup card but is a large load.
Typical local GPU	xtts, f5, chatterbox, outetts, csm, fish, bark, dia, orpheus, voxcpm2, vibevoice	NVIDIA GPU visible to WSL2 (nvidia-smi inside the distro). Estimates range from ~1 GB (kokoro/speecht5) to ~8 GB (dia/orpheus/voxcpm2).
Large GPU	higgs (~16 GB estimate, ~14.8 GiB observed), qwen (~16 GB estimate, ~7 GiB settled after a heavy load), voxtral (~20 GB estimate, ~19.5 GiB observed)	A high-VRAM NVIDIA card. Unload immediately after a batch.
Verification / captions	whisper	tiny/base are ~1 GB; large is ~10 GB. Used for SRT and optional verify-while-generating.

Default worker device is cuda:0 when nvidia-smi succeeds, otherwise cpu. Edge never uses the GPU. A machine with no NVIDIA GPU can still use Edge, VITS, SpeechT5, and (slowly) other engines on CPU; large GPU-first engines will be painful or will fail the reserve check.

9. A Hugging Face token for gated models

Many weight repos are on Hugging Face and some are gated. The Setup tab's HuggingFace Token field (hf_...) saves to E:\tts_server\secrets\hf_token (Linux path /mnt/e/tts_server/secrets/hf_token). Create a token at huggingface.co/settings/tokens. For fine-grained tokens, enable "Read access to contents of all public gated repos you can access". Classic Read tokens work out of the box. You must also accept each model's license on its Hugging Face page before the download will succeed.

CLI equivalents:

tts.cmd hf-token get
tts.cmd hf-token set hf_yourtokenhere

Empty string / empty Save removes the token. Edge TTS does not need a Hugging Face token. Local engines that pull gated or authenticated Hub files do.

10. Network for first-time installs and cloud Edge TTS

For optional engines, first-time Install on the Setup tab clones packages and downloads weights (Hugging Face Hub). That requires outbound HTTPS. Edge TTS is a cloud API: every Generate call needs internet, and the live voice catalog is fetched (cached six hours, with a static fallback of common en-US-JennyNeural-style names if the catalog request fails). After weights are on the VHDX, local engines can speak offline. Clearing the hub cache is dangerous for engines that treat Hub as their offline store; prefer tmp / pip when cleaning.

What you get once those requirements are met

• A WebView2 window titled Portable TTS Server V2 with tabs Setup, Server, Voices, Testing, Editor, Log, plus Shutdown.
• An authenticated HTTP API (token in X-TTS-API-Token, also accepted as Authorization: Bearer; the GUI's EventSource log stream uses ?token= because it cannot set a header).
• A Windows CLI at E:\tts_server\tts.cmd that auto-discovers the URL and token from the registry.
• Reference voices in E:\tts_server\voices, jobs in E:\tts_server\output\jobs, named projects in E:\tts_server\projects_output.

Related pages

• Start, stop, and portability
• Storage, layout, and ports
• Troubleshooting
• A to Z knowledge page