Voice AI is moving into a busy new phase, with open models targeting cloning, transcription, translation, and speech-based applications. On October 29, 2026, Edge0 released Audio8-TTS-Realistic, while Nvidia introduced Nemotron 3.5 Transcribe the day before.
The two releases take different paths. Audio8-TTS-Realistic creates speech from text and can clone a voice from one reference clip and its transcript. Nemotron 3.5 Transcribe turns recorded audio into text across as many as 40 languages. Together, they show how developers can now work with more parts of the speech pipeline using open-weight systems.
Edge0’s compact voice-cloning model
Audio8-TTS-Realistic is a 0.6B-parameter, open-weight text-to-speech model for zero-shot voice cloning. It supports 23 languages, which is 12 more than its predecessor, and its weights and code ship under the Apache 2.0 license. The model uses a 44.1 kHz neural audio codec and supports up to 2,048 packed text and audio positions of context.
Edge0 built the system around a dual-autoregressive, or DualAR, design inspired by Fish Audio S2 Pro. Its slow branch is based on Qwen2.5 0.5B, with 24 layers at width 896, and runs at about 21.5 frames per second. Each step predicts text tokens and the first audio codebook, while a shallow four-layer fast branch predicts the remaining nine codebooks in parallel.
That design keeps sequential computation at the frame level instead of repeating it for every audio codebook. The main model has 601 million parameters, excluding the codec. On an Apple M2 CPU with INT4 ONNX, peak memory use sits between 1.1 and 1.2 GiB.
Training runs in three stages. It includes multilingual supervised fine-tuning and GRPO, guided by a learned realism critic. The combined reward gives a 0.7 weight to WER and a 0.3 weight to SIM, while the critic compares each output with a matched human recording of the same sentence.
The critic learned from a quality ladder built with six open-source TTS systems. An identity-erasure loss keeps it from rewarding generator fingerprints, and its head has about 3 million parameters on top of a frozen WavLM-large encoder. A duration penalty also applies when generated speech passes 1.35 times the reference length.
What the comparisons show
In the final training stage, Edge0 used only Chinese and English prompts. Error fell from 22.53% to 20.77%, while SIM rose from 0.637 to 0.679. The published figures are self-reported error rates, where lower is better, and Edge0 calls the table a reference comparison because evaluation pipelines and text normalization differ.
Seed-TTS, tested in English and Chinese, records a 1.25% average error rate with a 0.704 SIM score. CV3, which covers nine languages, records 4.86% and 0.661. MiniMax, covering 24 languages, records 5.89% and 0.754.
On CV3, Audio8 beats VoxCPM2 at 5.35%, CosyVoice3-1.5B at 6.89%, and Qwen3-TTS-1.7B at 7.73%. On MiniMax, Qwen3-TTS-1.7B scores 27.41%, though it officially supports only 10 of the 24 test languages. Higgs TTS 3 reports 4.41% on CV3 and 2.74% on MiniMax, while Qwen3-TTS-1.7B averages 1.01% on Seed-TTS.
Developers need PyTorch 2.5+ and Transformers 4.57+ to load Audio8 with trust_remote_code=True. Voice cloning requires a clean reference clip and its exact transcript. The GitHub repository adds a command-line interface, batching, and supervised fine-tuning scripts, while an SGLang Omni adapter supports paged attention and dynamic batching for GPU serving.
Nvidia targets meetings, medicine, and edge hardware
Nvidia released Nemotron 3.5 Transcribe as a free, open-weight model for recorded audio. It can handle recordings up to an hour long, distinguish up to eight speakers, and translate 31 languages to and from English. One limitation is clear: Transcribe works with recorded audio, not livestreams.
Nvidia says the 30-billion-parameter model can run on a single chip, from the Jetson AGX Thor robotics controller and DGX Spark chip to Blackwell datacenter chips. The company also released version 2 of Granary, its open-source speech dataset, and points developers toward NeMo Speech for customizing systems to recognize industry jargon in each language.
The intended uses include business meetings and medical transcription. Nvidia also says Nemotron 3.5 beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard. The model joins Nvidia’s other open-weight voice systems, including Parakeet, which has 1.1 billion parameters.
A crowded speech and AI release cycle
The new models arrive during a busy stretch for voice technology. Mistral shipped Voxtral 3 earlier this month, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni.
ZenMux also added Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed from Moonshot AI. Kimi K3.1 Pro is listed as active alongside the Highspeed version, and Moonshot AI describes it as the first model in the K3.1 series and a post-trained version of Kimi K3. The developer reports a one-million-token context window, with improvements across software engineering, 3D creation, and research tasks.
Elsewhere, Genie 4 will roll out soon to developers and Google AI Ultra subscribers. A CNN report published on October 29, 2026, at 2:16 p.m. EDT also examined research showing that AI hardware can be hacked so chatbots provide dangerous answers.
That mix of releases points to a wider shift: speech models are becoming easier to run, customize, and connect to other AI systems, while researchers continue testing the hardware and safety limits around them.
Based on
- Edge0 Releases Audio8-TTS-Realistic, a 0.6B Voice Cloning TTS Model — marktechpost.com
- DeepMind’s Genie 4 world model adds audio and more realistic people to simulations | Ars OpenForum — arstechnica.com
- Voice AI in vogue: Nvidia releases new open transcription model | TechCrunch — techcrunch.com
- ZenMux Introduces Kimi K3.1 Pro for Long-Context AI Workflows – USA Today — usatoday.com
- AI hardware hacks can force dangerous chatbot answers — ctvnews.ca
