Open Source TTS and Audio Models: Maya 1, VibeVoice 7B, MMAudio V2 Compared (2026)
Last updated August 7, 2026
Three open audio models, three different licenses: Maya 1 (Apache 2.0, commercial OK — 3B voice-design TTS, 16GB GPU, sub-100ms), VibeVoice 7B (MIT via mirrors but Microsoft pulled the weights and discourages production use; ~45–90 min, 4 speakers, EN/CN only), and MMAudio V2 (video-to-audio; MIT code but CC-BY-NC weights — non-commercial). All three run in the invideo agent's Audio tab.
Updated August 2026
"Open source" covers three very different deals in AI audio, and the three open models in invideo's picker illustrate all of them. Maya 1 is genuinely open for business: Apache 2.0, commercial use allowed. VibeVoice 7B is MIT-licensed but vendor-discouraged: Microsoft pulled its original weights and labels the release research-only. MMAudio V2 is the trap for the unwary: MIT code, but CC-BY-NC weights — non-commercial, full stop. If you publish monetized work, the license column matters more than any benchmark, so this guide leads with it.
How do the three open audio models compare?
| Maya 1 | VibeVoice 7B | MMAudio V2 | |
|---|---|---|---|
| Maker | Maya Research | Microsoft | UIUC + Sony AI (CVPR 2025) |
| Job | Expressive voice design TTS | Long-form multi-speaker TTS | Video-to-audio (synced SFX/Foley) |
| Size | 3B params | ~9B actual (BF16) | Runs in ~6 GB VRAM |
| Output | 24 kHz, streaming | Up to ~45–90 min, up to 4 speakers | 44.1 kHz, ~8 s clips |
| Languages | English-focused | English + Chinese only | n/a (non-speech audio) |
| Code license | Apache 2.0 | MIT | MIT |
| Weights license | Apache 2.0 — commercial OK | MIT via community mirrors — vendor-discouraged for production | CC-BY-NC 4.0 — NON-commercial |
The pattern: "open source" tells you little until you check the weights license and vendor intent separately. All specs and license claims come from the projects' own model cards and repos, as of August 2026.
What is Maya 1?
Maya 1 is a 3B-parameter, Llama-style decoder-only TTS model whose pitch is voice design rather than voice selection — per its Hugging Face model card: "Create any voice you can imagine — a 20s British girl, an American guy, or a full-blown demon." You condition it with a natural-language description in an XML attribute (<description="40-year-old, warm, low pitch, conversational">) and steer delivery with 20+ inline emotion tags — laugh, cry, whisper, rage, sigh, gasp.
It emits SNAC neural-codec tokens at 24 kHz (~0.98 kbps), streams at sub-100 ms latency with vLLM, and runs on a single 16 GB+ GPU (RTX 4090 class). License: Apache 2.0 — game characters, podcasts, and monetized video are all fair use. Caveat: benchmark claims are self-reported, with no independent leaderboard placement as of August 2026.
What is VibeVoice 7B, and why is its status complicated?
VibeVoice is Microsoft's long-form conversational TTS family — the podcast model. The 7B/Large variant (~9B actual parameters on a Qwen2.5 base with a diffusion decoding head) generates up to roughly 45 minutes in one pass within a 32K context, with the family documented "up to 90 minutes… up to 4 distinct speakers." For multi-host podcast-style audio, nothing else open comes close on length.
Now the complication. Microsoft removed the original microsoft/VibeVoice-Large weights from Hugging Face — the official repo now returns an access error — and the 7B weights circulate through community mirrors under the original MIT license. Meanwhile Microsoft's own responsible-AI notes call the release "intended for research and development purposes only," flag impersonation and disinformation risks, and recommend against commercial deployment without further testing. The legal position (MIT) and the vendor's position (research-only) point in different directions. Documented hard limits: English and Chinese only — other languages produce unexpected output — and speech only: it "does not handle background noise, music, or other sound effects," per the model card.
What is MMAudio V2 — and can you use it commercially?
MMAudio V2 (large_44k_v2) is the odd one out — not TTS, but video-to-audio. From UIUC and Sony AI, published at CVPR 2025, it "generates synchronized audio given video and/or text inputs" — Foley and sound effects timed to on-screen action. The flagship weights output 44.1 kHz audio in ~8-second (variable) clips and run in about 6 GB of VRAM at 16-bit.
The license catch, prominently: the code is MIT, but the pretrained weights are CC-BY-NC 4.0 — non-commercial. The authors state it directly: "We do not guarantee that the pre-trained models are suitable for commercial use." For monetized content, self-hosted MMAudio weights are off the table.
The authors are unusually frank about quirks — documented failure modes include occasionally emitting "unintelligible human speech-like sounds," adding unsolicited background music, and struggling with unfamiliar concepts.
Which open model should you actually pick?
- Commercial voice content on your own hardware → Maya 1: license, vendor intent, and capability all align.
- Long multi-speaker episodes for research or internal use → VibeVoice 7B, eyes open about the pulled weights and the EN/CN limit.
- Synced SFX for video → MMAudio V2 for non-commercial work; monetized video needs a commercially licensed pipeline.
What can you make with them?
Voice-design models like Maya 1 map to the character voiceover and narration an AI voice generator workflow handles end to end; video-to-audio like MMAudio is the research cousin of the add audio to video path, where licensing is already sorted for commercial use.
Open-model licensing questions, answered
Can I use Maya 1 commercially?
Yes. Maya 1 is Apache 2.0 — code and weights — permitting commercial use, modification, and redistribution: the cleanest license of the three.
Can I use VibeVoice 7B commercially?
Legally, the MIT license on the mirrored weights permits it. Practically, Microsoft pulled the originals and labels the model research-only, recommending against production use without further testing. Most commercial teams should treat that as a no.
Can I use MMAudio V2 commercially?
No — not the pretrained weights. They're CC-BY-NC 4.0 (non-commercial), even though the code is MIT, and the authors explicitly decline to guarantee commercial suitability.
What hardware do these models need?
Maya 1 runs on a single 16 GB+ GPU with sub-100 ms streaming via vLLM; MMAudio V2 needs about 6 GB of VRAM at 16-bit; VibeVoice 7B is the heaviest at ~9B actual parameters in BF16.
What languages do the open TTS models cover?
Far fewer than commercial engines: VibeVoice supports only English and Chinese, and Maya 1 is English-focused. Broad multilingual coverage remains a paid-API advantage.
Why did Microsoft pull the VibeVoice weights?
Microsoft hasn't published a detailed explanation; its responsible-AI notes cite impersonation and disinformation risk and research-only intent. The narrative is partly community-reconstructed — the verifiable facts are the inaccessible official repo and the MIT-licensed mirrors.
Where can you try all three?
Self-hosting is the open-source promise — and also GPUs, codecs, and license reading. The shortcut: Maya 1, VibeVoice 7B, and MMAudio V2 are all runnable from the Audio tab of the invideo agent, in the same picker as the commercial voice engines — so you can compare an open model against a paid one on the same script before standardizing. The complete roster is on the invideo AI models index.
Version history: MMAudio published at CVPR 2025 (flagship large_44k_v2 weights); VibeVoice family released 2025, original Microsoft weights subsequently pulled, 7B mirrors persist under MIT; Maya 1 released on Hugging Face under Apache 2.0. All license and spec claims checked August 2026 against each project's official model card or repository.