QVAC OpenMOSS Audio Demo

Three open audio models from OpenMOSS, running on your own computer with the QVAC SDK: a voice that reads any line in 31 languages, a dialogue between up to five voices, sound effects from a sentence, and a sound scene that puts a generated ambience under a dialogue. Nothing is sent anywhere: the models download once, then everything runs locally.

This is an example, not a product. A self-contained prototype showing what the QVAC SDK can do with OpenMOSS, provided as is, with no support and no warranty.

What it does

Tab	Model	What you can do
Expressive speech	MOSS-TTS v1.5	Read a line in 31 languages; clone a voice from one recording, no transcript needed; [pause 2.0s] markers; IPA (/təˈmɑːtoʊ/) or Pinyin (ni3 hao3) for pronunciation; a target length; the audio plays while the rest is still being generated
Dialogue	MOSS-TTSD	A script tagged [S1] to [S5], voiced in one pass, with the model's voices or with voices you record (each speaker reads one sentence aloud)
Sound effects	MOSS-SoundEffect v2	Describe a sound, pick 4 to 30 s and a quality, list what to avoid; 48 kHz
Sound scene	MOSS-TTSD + MOSS-SoundEffect	A dialogue over a 30 s ambience (a dog barking, rain, an office), looped if needed and mixed 10 dB under the voices; the same ambience is reused by the next take

Every take is saved in out/ with its settings, and "Another take" runs the same settings with a new random seed. The examples on each tab keep the seed they were picked with, so they play back the same way.

Quickstart

npm install
npm start

Open http://localhost:3102. Each tab says what it still needs to download; the button downloads it from the QVAC Registry with progress. After that, the demo works offline.

The server only answers this computer. To open it from a phone on the same network, start it with HOST=0.0.0.0 ALLOWED_HOSTS=<this computer's address>:3102 npm start; anyone on that network can then use it. Remove deletes a take's audio and a recorded voice from the disk, not only from the list.

The models

All five files come from the QVAC Registry and land in the SDK's model cache, ~/.qvac/models.

File	Used by	Size
MOSS-TTS v1.5 backbone (moss-tts-delay-f16.gguf)	Expressive speech	17.0 GB
MOSS-TTSD backbone (moss-ttsd-f16.gguf)	Dialogue, sound scene	16.7 GB
MOSS codec decoder (moss-codec-decoder-f16.gguf)	Speech, dialogue, scene	1.8 GB
MOSS codec encoder (moss-codec-encoder-f16.gguf)	Cloning a voice	1.8 GB
MOSS-SoundEffect v2 (moss-sfx-v2-q8_0.gguf)	Sound effects, sound scene	3.5 GB

40.8 GB for everything. Each tab downloads only what it uses: 20.5 GB for speech, 20.3 GB for the dialogue, 3.5 GB for sound effects.

Measured

On a 36 GB MacBook Pro (Apple M5 Max, Metal), QVAC SDK 0.21.0:

Job	Audio	Time to make
Expressive speech, one line	9.4 s	6.6 s, first sound after 2.6 s
Cloned voice, one line	5.8 s	4.7 s
Dialogue, three model voices	10.6 s	6.4 s
Dialogue, two recorded voices	9.0 s	6.3 s
Sound effect, Draft (25 steps)	4.0 s	24 s
Sound effect, Standard (50 steps)	8.0 s	48.5 s
Sound effect, Full (100 steps)	8.0 s	97.6 s
Sound scene, new Standard ambience	12.8 s	55.7 s

Loading a speech backbone takes about 5 s. A sound effect costs the same at any length: the model always works on a 30 s clip and crops it, so the quality setting (the number of steps) is what sets the time.

How it is built

• server.js: a node:http server bound to 127.0.0.1. Speech and dialogue go through the QVAC SDK (loadModel with ttsEngine: 'moss', then textToSpeech with stream: true). One speech model is loaded at a time: the two backbones are 17 GB each.
• engine/sfx.cjs: one sound effect in the Bare runtime, then exit. The sound-effect engine is in @qvac/tts-ggml from 0.11.0, a Bare addon, and the SDK 0.21 does not expose it yet.
• lib/models.mjs: the five registry files, where the SDK caches them, and the downloads.
• lib/wav.mjs: WAV reading and writing, and the scene mix.
• public/: the page. It plays speech while it is generated, records the voices to clone and converts any audio file to the 24 kHz mono WAV that MOSS clones from.

Known limits

• A dialogue sometimes does not end. With the SDK 0.21 defaults, MOSS-TTSD sometimes speaks the lines and then keeps generating silence up to its 162.6 s budget, and with some seeds a take with recorded voices is silent from the start. The demo watches the audio as it arrives: 5 s of silence after the last word ends the take there, and no speech in the first 8 s stops it with a message. "Another take" then tries another seed. In our runs with recorded voices, three seeds out of four gave a take, and one of the three dropped the end of its last line.
• Sound effects come from @qvac/tts-ggml 0.11, not from the SDK yet. The SDK 0.21 bundles 0.10, which has no sound-effect engine, so the demo depends on 0.11 directly and runs it in Bare. When an SDK release exposes MOSS-SoundEffect, that tab can move to the SDK like the others.
• Dialogue languages. The script tags and the examples are English; MOSS-TTSD is multilingual, and the language picker sets its hint.

Recommended hardware

	Requirement
Computer	Tested on macOS with Apple Silicon. MOSS runs its 8B backbones on a desktop GPU
RAM	32 GB or more (tested on 36 GB). A speech backbone with its codec is about 21 GB in memory
Disk	41 GB for the five models, plus the SDK
Microphone	Only to record a voice to clone; a file works too

Requirements

• Node.js 22.17 or newer
• About 45 GB of free disk

Licences

The demo is Apache 2.0. MOSS-TTS v1.5, MOSS-TTSD and MOSS-SoundEffect are Apache 2.0 models by OpenMOSS (Hugging Face); the weights download from the QVAC Registry and are not in this repository.