VoxLoom: A Free Alternative to ElevenLabs and WisprFlow

If you have looked at AI voice tooling recently, you have probably seen the same two names everywhere. ElevenLabs owns speech output — realistic text-to-speech and voice cloning, billed per character in the cloud. WisprFlow owns speech input — AI dictation that types what you say, at $144 a year, also in the cloud. Between them they cover the whole voice I/O loop, and both of them sit on somebody else’s servers.

VoxLoom is a free, local-first alternative that does both halves in one app — and runs the entire loop locally on your machine.

The short version

	VoxLoom	ElevenLabs	WisprFlow
Voice cloning	✅ Local, zero-shot	✅ Cloud	❌
Text-to-speech	✅ 7 engines, 23 languages	✅ Cloud	❌
System-wide dictation	✅	❌	✅ Cloud
AI agent voices (MCP)	✅	❌	❌
API	✅ Local REST + MCP, no keys	✅ Metered, key required	❌
Runs offline	✅	❌	❌
Price	Free (limited time)	$5–330/month	$144/year

The trade-off is simple: you need your own hardware (an Apple Silicon Mac, an NVIDIA GPU, or a ROCm/Arc card), and you install an app instead of creating an account. In exchange, generation is unmetered, your voice data never leaves the machine, and you get an API with no rate limits.

Why look beyond per-character pricing

The economics of cloud voice services bite hardest for heavy users. A single YouTube narration can run 800–1,500 words. Audiobooks, podcasts, game dialogue, and localization multiply that quickly, and free tiers evaporate. Meanwhile, dictation subscriptions add a second recurring bill on top.

There is also the privacy question. Your voice is one of the most personal identifiers you own. Cloud pipelines ship your reference audio, your scripts, and your rough spoken drafts to third-party servers. For lawyers, healthcare workers, and anyone handling sensitive material, that is a hard no — and “we don’t train on your data” is a policy that can change, not an architectural guarantee.

Local-first tools flip the default: the data path physically never touches the network. Privacy stops being a promise and becomes a property of the system.

What VoxLoom actually is

VoxLoom is a local-first AI voice studio built with Tauri (Rust) — a native desktop app, not an Electron wrapper. It bundles the full voice stack into one installer:

• Zero-shot voice cloning — clone a voice from a few seconds of reference audio, no fine-tuning
• 7 TTS engines in 23 languages — from English to Arabic, Japanese, Hindi, and Swahili
• System-wide dictation — a global hotkey that works in any app: hold, speak, release
• Agent voices — a built-in MCP server so AI agents can speak in voices you own
• Stories editor — a multi-track timeline for podcasts, conversations, and narratives
• Audio effects — pitch shift, reverb, delay, chorus, compression, and filters
• Local REST API — every engine becomes an endpoint on localhost, no keys, no quotas

The seven engines

Most open-source voice tools give you one model and a Python environment. VoxLoom ships seven engines behind one interface, so you can pick per project:

Engine	Best for
Qwen3-TTS (0.6B / 1.7B)	High-quality multilingual cloning with voice instructions
Qwen CustomVoice	50+ preset voices with natural-language delivery control
Chatterbox Multilingual	Conversational multilingual speech
Chatterbox Turbo	Expressive paralinguistic tags — [laugh], [sigh], [gasp]
LuxTTS	Ultra-lightweight (~1GB VRAM), 48kHz, fast on CPU
HumeAI TADA	Emotionally expressive delivery
Kokoro	Compact, fast preset voices

Speech-to-text runs on Whisper in five sizes (base through large, plus Turbo), with a bundled Qwen3 1.7B local LLM that refines raw transcripts — stripping filler words and self-corrections without rephrasing your meaning.

Voice cloning, locally

Cloning is zero-shot: record or drop in a reference clip, and the voice is added to your library immediately. Because inference happens on your GPU (Metal on macOS, CUDA on Windows and Linux, ROCm for AMD, plus Intel Arc), it stays responsive and works fully offline. A cloned voice works everywhere in the app — generation, dictation replies, and agent speech.

Dictation into any app

This is the WisprFlow half. Press and hold a global hotkey, speak, release — your words land in whatever field is focused, in any application. Push-to-talk and toggle modes are both supported, auto-paste is accessibility-verified on macOS, and every capture keeps its original audio alongside raw and refined transcripts, archived and searchable.

Give your agents a voice

The most distinctive feature is the built-in MCP server. One tool call — voxloom.speak — and an MCP-aware agent such as Claude Code, Cursor, or Cline speaks its answer aloud in a voice you’ve cloned. Bind each client to its own voice profile and you can tell which agent is talking without looking at the screen. Every agent-initiated speech surfaces a visible pill, so nothing speaks silently in the background.

Voice profiles can also carry personas — free-form character descriptions processed by the bundled local LLM. Compose a fresh line in character, rewrite your text in their voice, or have the persona respond to a message; agents can invoke the same modes over MCP.

Who it is for

• Content creators — audiobook narration, YouTube voiceovers, podcast intros. Unlimited length with auto-chunking and crossfade, in any voice you own the rights to.
• Developers — a localhost REST API with no keys and no rate limits. Wire voice into apps, games, and scripts; generate NPC dialogue on the fly; or replace a cloud TTS line item entirely.
• AI power users — agents that talk back, per-agent voice profiles, and persona-driven generation, all local.
• Privacy-sensitive teams — legal, healthcare, finance: the audio never leaves the device, full stop.

Getting started

1. Download the installer for your platform — macOS (Apple Silicon), macOS (Intel), Windows MSI, or run docker compose up for a containerized setup.
2. Pick your first engine. Kokoro and LuxTTS are light and run almost anywhere; Qwen3-TTS is the quality ceiling.
3. Drop in a few seconds of reference audio to clone a voice, or start with the 50+ preset voices.

Honest limitations

No tool is right for everyone, and VoxLoom is no exception:

• Hardware matters. The top-tier engines want a real GPU. If you are on an older CPU-only machine, stick to Kokoro, LuxTTS, and the smaller Whisper models.
• Setup is an install, not a signup. You download models (multi-GB for the large engines) on first use — after that, everything is offline.
• License diligence. VoxLoom is free for a limited time, but individual engines carry their own licenses. Check the engine’s terms before commercial use — same as with any TTS stack.
• Cloud conveniences. If you need studio-grade dubbing workflows or a browser tool on a locked-down machine, a hosted service may still fit better.

For everything else — unmetered generation, private cloning, dictation, and agents that speak — VoxLoom covers the whole loop for the price of your existing hardware.