Three Local Voice Models Land Under One Mirror Account

What happened

NostraEmpire published three local-weight voice models to Hugging Face on 31 August 2026: Parler-TTS Large v1 (2.2B, Apache-2.0), Qwen2.5-Omni-7B (7B, licence tagged "other"), and MeloTTS-French (MIT). All three ship as safetensors and load through the transformers library.

Context

None of the three models are new. Parler-TTS is a reproduction of work by Dan Lyth and Simon King at Stability AI and Edinburgh University (arXiv 2402.01912). Qwen2.5-Omni is an end-to-end multimodal release with streaming speech output. MeloTTS is MyShell.ai's VITS-based multi-lingual TTS library. What is new is the timing: three models covering lightweight CPU TTS, prompt-controlled GPU TTS, and streaming multimodal voice, published the same day under a single account.

How it works

Parler-TTS Large v1 is a 2.2B-parameter model that takes a natural-language prompt to control speech features: gender, background noise, speaking rate, pitch, and reverberation. It was trained on 45K hours of audio across 34 named speakers, so you select a consistent voice by name. Inference runs on CUDA or CPU via the transformers pipeline. The model card references SDPA, torch.compile, batching, and streaming for optimisation but gives no VRAM figure.

Qwen2.5-Omni-7B uses a Thinker-Talker architecture. The Thinker component processes text, images, audio, and video; the Talker generates text and natural speech in a streaming fashion. TMRoPE (Time-aligned Multimodal RoPE) synchronises video timestamps with audio. The model supports chunked input with immediate output for real-time voice and video chat and scored 56.13% average on OmniBench.

MeloTTS-French is the lightest of the three. Built on VITS and Bert-VITS2 from the Coqui TTS project, it handles French, multiple English variants, Spanish, Chinese, Japanese, and Korean, and is fast enough for CPU real-time inference.

Our read

The interesting story is not any single model; it is the combination and the provenance question. For a studio running ComfyUI on local hardware, these three slots cover a full voice pipeline: MeloTTS for quick multilingual VO on a CPU, Parler-TTS for prompt-controlled narration ("female voice, slight reverb, 15% slower" without a reference clip), and Qwen2.5-Omni for a single model that ingests video and audio and speaks back in streaming fashion. That last one collapses a video-understanding-plus-TTS pipeline into one model.

The friction is the mirror account. NostraEmpire is not the upstream publisher for any of the three. No source states whether this is an official mirror, a community re-upload, or an automated sync, and no update or deprecation policy is given. A production workflow should pin to the upstream repositories rather than the mirror handle.

The second friction is Qwen2.5-Omni's licence. The hub tag reads "license:other" and the model card names no licence. Until upstream terms are confirmed, shipping a video product that calls this model commercially carries a legal risk the other two do not. Parler-TTS (Apache-2.0) and MeloTTS (MIT) are unambiguous.

What this changes

If your current TTS step in ComfyUI is an API call, Parler-TTS Large v1 is a local replacement: load it via transformers, describe the voice in plain English, and stop paying per character. For French or multilingual quick VO, MeloTTS runs on CPU in real time and needs no GPU slot. Qwen2.5-Omni-7B is the one to watch and the one to hold: at 7B in bf16 the weights are roughly 14 GB plus KV-cache overhead, so an 8 GB card is unlikely to suffice. If you have 24 GB or more and can confirm the upstream Qwen licence permits commercial use, it is the most operationally interesting model here. Until that licence question is resolved, the practical Monday change is Parler-TTS and MeloTTS.

License

Parler-TTS Large v1: Apache-2.0. Commercial use, modification, and redistribution are permitted; include the licence and attribution notice.

MeloTTS: MIT. Free for commercial and non-commercial use with attribution.

Qwen2.5-Omni-7B: the hub tag is "license:other" and the model card names no specific licence. The sources do not state the terms. Check the upstream Qwen repository before building anything commercial on it.

Key takeaways

• Three local-weight voice models landed on Hugging Face under one mirror account on 31 August 2026, covering CPU TTS, GPU prompt-controlled TTS, and streaming multimodal voice.
• Parler-TTS Large v1 (2.2B, Apache-2.0) controls speech characteristics via a text prompt and is clear for commercial video work.
• MeloTTS-French (MIT) runs real-time on CPU across six language families; it is the zero-GPU option for multilingual VO.
• Qwen2.5-Omni-7B is the most capable for video-plus-voice workflows, but its commercial licence is unconfirmed in the sources.
• All three are mirrors under a single account with no stated update policy; pin to upstream repositories for production.

Sources

1. NostraEmpire/mirror-parler-tts-large-v1 — text-to-speech on Hugging Face — tier 3
2. NostraEmpire/mirror-qwen2.5-omni-7b — any-to-any on Hugging Face — tier 3
3. NostraEmpire/mirror-melotts-french — text-to-speech on Hugging Face — tier 3