GLOAM.FM
VOICE STUDIO
Features Models Intelligence Voice Lab Developers Docs Changelog Pricing Download
100% ON-DEVICE · NO CLOUD

Voice cloning
that never
phones home.

Native macOS voice studio for Apple Silicon. Design a voice, direct its emotion, and voice your Gloam.fm AI DJ — every sample rendered locally with MLX.

Download for macOS ★ Star on GitHub
Free · MIT licensed · macOS 14+ · Apple Silicon
7+
SYNTHESIS MODELS
~1–3×
RTF BY BACKEND
0
BYTES UPLOADED
MIT
OPEN SOURCE
GLOAM.FM DJS
Voice your station
Clone the hosts and MCs that drive your AI radio, then run them live over the API.
PODCASTERS
Script & narrate
Draft in Voice Lab, direct each line's emotion, and export one stitched, normalized WAV.
GAME & APP DEVS
Local voice pipeline
Batch-generate character lines against a private, OpenAI-compatible endpoint — no per-call fees.
ACCESSIBILITY
A private voice
Bank and reuse a personal voice entirely offline — nothing ever uploaded to a server.
EVERYTHING, ON YOUR MAC

A complete voice studio

01

Voice cloning

Drop a reference clip and clone a voice in seconds. Save, edit and export as portable .gvoice packs.

02

Emotion & speed

Five emotion variants — flat, neutral, warm, excited, hype — plus 0.5×–2.0× playback speed control.

03

A/B variants

Generate two takes side-by-side and compare waveforms and playback before you commit.

04

History & reuse

Every take is saved with backend, voice, emotion and RTF. Play, delete, or one-click reuse into the editor.

05

On-device speech-to-text

Reference clips auto-transcribe. Dictate anywhere, or convert any audio file with ⇧⌘T — Apple or Whisper, always local.

06

In-app model downloads

Download weights with live progress and an automatic storage preflight. No terminal required.

PICK YOUR BACKEND

Seven synthesis models, one studio

Swap engines per take. From real-time lightweight cloning to premium research-grade quality and multilingual voice design — all downloaded and run on your Mac.

Chatterbox

MIT

Fast, lightweight real-time cloning. The everyday workhorse for quick iterations and live-feel synthesis.

RTF
~2–3×
QUALITY
Good
BEST FOR
Real-time

Chatterbox-Turbo

MIT

The faster sibling of Chatterbox — RTF ~1–2×. Trades Chatterbox's exaggeration knob for a quicker, more balanced take.

RTF
~1–2×
QUALITY
Higher
BEST FOR
Balance

Fish S2-Pro

RESEARCH

Premium research-grade quality. Per-session temperature control for expressive, nuanced delivery.

RTF
~1×
QUALITY
Premium
BEST FOR
Mastering

Qwen3-TTS family

APACHE-2.0

Multilingual and commercial-friendly. Clone at two sizes or direct a preset speaker — or invent a voice from a text prompt in Voice Design.

Chatterbox & Turbo weights are MIT · Qwen3 Apache-2.0 · Fish S2-Pro under the Fish Audio Research License
HEAR THE RANGE

Five emotions, one voice

Same speaker, same line — steered from flat to hype. Tap an emotion to hear how the delivery changes.

“You're locked in to Gloam FM — let's take it up a notch.”
now playing · warm
Samples render on-device · drop your own clips in samples/
QWEN3-DESIGNNO REFERENCE NEEDED

Describe a voice. Hear it.

No clip to clone from — just write what you want and Qwen3-design conjures a speaker to match, entirely on-device.

VOICE DESCRIPTION
designed · nightowl
Rendered on-device by qwen3-design · sample plays from samples/qwen3-design.wav
ON-DEVICE INTELLIGENCEPREVIEW

Local LLMs, ready to write

A built-in text-model catalog runs Gemma-4 and Qwen3 entirely on your Mac — the engine that will script DJ banter, station IDs and show intros. Pick a tier to match your RAM; heavier weights, sharper writing.

MODEL
FAMILY
SIZE
TIER
gemma4-e2b
Gemma-4
1.6 GB
Everyday
gemma4-e4b
Gemma-4
3.0 GB
Sharper
gemma4-26b
Gemma-4
15 GB
Genius · MoE
gemma4-31b
Gemma-4
18 GB
Let it rip · 64GB+
qwen3-1.7b
Qwen3
1.1 GB
Lightweight
qwen3-8b
Qwen3
4.7 GB
Balanced
MoE routing for Gemma-4 26B (4B active)
Thinking-off, streaming & tool calls
4-bit MLX conversions, downloaded in-app
MATCH YOUR MAC
16 GBEVERYDAY
Runs the e2b/e4b Gemma-4 and Qwen3-1.7B tiers comfortably alongside voice synthesis.
32 GBSHARPER
Adds headroom for Qwen3-8B and the Gemma-4 26B MoE (4B active) “Genius” tier.
64 GB+LET IT RIP
Unlocks the dense Gemma-4 31B for the sharpest writing the catalog offers.
VOICE LAB

Write a script. Direct every line.

A multi-line editor where each line carries its own voice and emotion direction. Reorder freely, generate the whole session in one click, and export a single stitched WAV.

Persistent takes
Every take saved to disk (WAV + JSON) and browsable per line — survives relaunches.
Batch generation
Render an entire script serially with per-line status: queued → generating → done.
Stitched export
One WAV, configurable silence gaps, optional peak normalization to −0.18 dBFS.
Direction overrides
Per-session temperature and exaggeration knobs that override emotion presets.
Voice Lab — Script
[ app screenshot — Voice Lab ]
MADE FOR GLOAM.FM

Give your AI DJ a voice

Design and clone the hosts, MCs and DJs that drive your Gloam.fm station. Export them as .gvoice packs and drive them live through the local API — the same interchange format the Python engine speaks.

Explore Gloam.fm →
FOR DEVELOPERS

A local, OpenAI-compatible API

Flip on the built-in HTTP server and drive synthesis from any script. Loopback-only, port 8790, minimal sandbox entitlements — nothing exposed to the network.

Drop-in /v1/audio/speech endpoint
Drives (and is driven by) the Python gloam-voice-engine
Portable .gvoice packs shared across both projects
Read the docs →
zsh — gloam
$ curl localhost:8790/v1/audio/speech \
-d '{"voice":"my-voice",
"input":"You're locked in to Gloam FM.",
"emotion":"hype"}' -o out.wav
→ 200 OK · audio/wav · 1.4s
→ nothing left the device
PRIVATE BY ARCHITECTURE

Your audio never leaves your Mac

No accounts

No sign-in, no sync, no telemetry. Just open the app and go.

Sandboxed

Minimal entitlements; data lives in Application Support and Caches only.

Loopback API

The optional server binds to localhost only — never the open network.

Yours to export

Voices and takes are plain files you own — .gvoice, WAV, JSON.

ON-DEVICE VS. THE CLOUD

Why local wins

Gloam Voice Studio
CLOUD TTS
Runs fully offline
Always
Never
Your audio stays local
On-device
Uploaded to servers
Account required
None
Sign-up + API key
Cost
Free · MIT
Per-character / sub
Generations
Unlimited
Metered / quota
Voice ownership
Portable files
Locked to platform
Works on a plane
Yes
No
Comparison reflects typical hosted TTS services · your mileage may vary by provider
WHERE IT'S HEADED

Roadmap

SHIPPING NOW
7 synthesis backends
Voice Lab + stitched export
Local OpenAI-compatible API
On-device speech-to-text
IN PREVIEW
On-device LLM catalog (Gemma-4 / Qwen3)
MoE routing for large Gemma-4
Direction overrides per session
NEXT
LLM-authored DJ banter in-app
Mac App Store release
More multilingual voices
Deeper Gloam.fm live integration
Roadmap is indicative, not a commitment · follow along on GitHub
UP AND RUNNING IN MINUTES

Three steps to your first take

1

Download & open

Grab the app or build from source. macOS 14+, Apple Silicon. Pick a model and it downloads in-app.

2

Add a reference

Record or drop a short clip. It auto-transcribes on-device; an optional transcript hint sharpens quality.

3

Generate

Type your line, choose emotion and speed, and synthesize. Compare A/B takes and export.

QUESTIONS

Frequently asked

Is it really free?
Yes. The app is MIT licensed and free forever — download it, build it from source, or fork it. There are no accounts, subscriptions, or paid tiers. Sponsorship is optional.
Which Macs are supported?
Apple Silicon only (M1, M2, M3, M4) running macOS 14 Sonoma or newer. Intel Macs are not supported because synthesis runs on the Apple Silicon GPU via MLX.
Do I need an internet connection?
Only once, to download the model weights you want. After that everything — cloning, synthesis, speech-to-text and the local LLMs — runs fully offline. No audio ever leaves your Mac.
How big are the downloads?
Voice models range from roughly 1–4 GB. The optional on-device text models (Gemma-4 / Qwen3) range from ~1 GB up to ~18 GB depending on the tier you pick. A storage preflight checks space before downloading.
Can I use generated voices commercially?
It depends on the model weights. Chatterbox and Turbo are MIT, and Qwen3 is Apache-2.0 (commercial use permitted). Fish S2-Pro is under the Fish Audio Research License — personal/research use only unless you license it from Fish Audio.
What is the difference between the voice models and the LLMs?
The seven synthesis models turn text into speech. The on-device intelligence catalog (Gemma-4 / Qwen3 text models) is a separate, preview feature that generates the words — DJ banter, station IDs, show intros — that you then voice.

Start cloning locally

Free and open source. Nothing to sign up for — download and generate your first voice in minutes.

↓ Download for macOS ★ Star on GitHub
Requires macOS 14+ · Apple Silicon (M-series)
STANDING ON SHOULDERS

Built on open source

Gloam Voice Studio is made by Tiny Trash Labs on a stack of remarkable open projects. Thanks to everyone who builds and maintains them.