Search “Speech”

Edge0 Releases Audio8-TTS-Realistic: A 0.6B Open-Source Voice Cloning TTS Model

Edge0 has released the open-weight voice cloning TTS model Audio8-TTS-Realistic, with 601M parameters, supporting 23 languages, with weights and code released under Apache 2.0. The model requires only 1 reference audio clip and its corresponding text for zero-shot voice cloning, and can also directly generate speech when no reference voice is available. Self-reported results show an average error rate of 4.86% on the CV3 nine-language benchmark, lower than Qwen3-TTS-1.7B's 7.73%, but it lags behind 4B-class models on the MiniMax benchmark, with Thai WER reaching 48.57%.

10-person team Phonic releases speech model Phonic-1, claims it can query and book in parallel during calls

The 10-person startup Phonic has released a speech model, Phonic-1, which it says can execute queries and bookings before the caller has finished speaking. The model tops the Artificial Analysis speech-to-speech leaderboard with a score of 87.1, 0.2 points higher than xAI's Grok Voice 3.0, with a customer service task completion rate of 79.9% and a first-audio latency of 0.61 seconds. Phonic says it scored 90.2 on the Cekura call test set, but the public Cekura leaderboard lists Phonic v1 at 68.3%; the company has completed a $27.5 million Series A round led by First Round, bringing total funding to $31.5 million.

NVIDIA Releases Nemotron 3.5 Transcribe Open-Weight Speech Model

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight 30B MoE model with approximately 4B active parameters. The model supports transcription in 40 languages, speaker attribution, translation, contextual prompting, and audio understanding, and ranks first in VoiceArena's 14-system comparison of speaker diarization plus ASR.

NVIDIA Releases Nemotron 3.5 Transcribe Open-Weight Speech Model, Supporting 40 Languages

On October 28, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that can transcribe audio in 40 languages, translate speech, and identify speakers, and developers can control the transcription format with natural language instructions without retraining. The model consists of a Transformer speech encoder, a speaker diarization model, and the Nemotron 3.5 Lightning language model, with approximately 4.3 billion activated parameters, and the released checkpoint is labeled 30B-A4B. NVIDIA says it achieves an average word error rate of 4.71% on nine public test sets of the English Open ASR Leaderboard and 4.62% on seven languages of the multilingual leaderboard, and calls it the best open-weight model on both leaderboards; the model is released in BF16, FP8, and NVFP4 formats as well as GGUF files, under the OpenMDW-1.1 license.

NVIDIA Releases Speech Transcription and Translation Model Nemotron 3.5 Transcribe

On October 28, 2026, NVIDIA released the speech transcription and translation model Nemotron 3.5 Transcribe, supporting 40 languages including Japanese and capable of transcribing recordings up to 1 hour long. The model is an open-weight 30B MoE architecture that activates approximately 4B parameters during inference, can specify via text instructions whether to retain filler words and how numbers are written, and supports speaker diarization for up to 8 people and bidirectional Japanese-English translation. The weights are released on Hugging Face as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B under the OpenMDW-1.1 license, with BF16, FP8, and NVFP4 versions and NeMo Speech fine-tuning tools provided.

NVIDIA Releases Nemotron 3.5 Transcribe: An Open-Source 30B MoE Speech Model…

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight large speech model that supports 40 languages and up to 1 hour of audio per pass, and can select output formats such as verbatim, written, speaker-labeled, contextual, and translation via text instructions. The model has approximately 34B total parameters and about 4.3B activated parameters, adopts a hybrid Mamba-Transformer MoE architecture, can run in BF16 on a single H100 80GB, and the NVFP4 version at about 21GB can run on an RTX 5090. NVIDIA self-reports an average WER of 4.71% across 9 public English Open ASR test sets, the best result among open-weight models, though commercial APIs still lead; its shortcomings are that it only supports offline operation, and the Chinese far-field meeting cpCER is 34.95%.

Block Releases Dessa, a Fully Duplex Speech Model Trained Purely on Speech

Block's AI research lab has released Dessa, a full-duplex speech model, opening a waitlist in the form of a research preview. It is trained only on human voice conversation audio, without using text or pretrained models. The model can run locally on a MacBook Pro M4 Max or higher configuration, faster than real time; training used one NVIDIA GB200 NVL72 (72 GPUs). No benchmark scores, parameter count, weights, paper, or API have been published so far.

NVIDIA releases the Nemotron 3.5 Transcribe speech transcription model, tying for first among open-source models with 2.8% WER

NVIDIA has released the open-weight speech transcription model Nemotron 3.5 Transcribe, with 30B total parameters and 4B activated parameters. On Artificial Analysis's non-streaming speech transcription leaderboard, it achieves a 2.8% AA-WER, tying for first among open-source models with Mistral's Voxtral Small. On the speech agent dialogue dataset AA-AgentTalk, it leads Voxtral Small's 2.3% with 1.8%. Its activated parameters are 4B, while Voxtral Small's are 24B; other open-source models of comparable scale have a WER of about 4%.

Block releases Dessa, a full-duplex model trained purely on speech, which can run locally in real time on a MacBook Pro M4 Max

Block, founded by Twitter co-founder Jack, has released Dessa, a speech-to-speech model trained from scratch purely on human speech, with no text and no pretrained base, supporting full duplex so it can listen and speak at the same time. The model claims inference faster than real time and can run locally on a MacBook Pro M4 Max and above, and the entire model is said to have been trained on a single NVIDIA GB200 NVL72. The team says its goal is to let the model understand tone and emotion, avoiding the safety risk of continuing to respond without perceiving the other person's distress.

Block launches full-duplex speech model Dessa, supports local operation

Block has launched the full-duplex speech model Dessa, which can listen to the user's speech while speaking, and can simulate emotional expressions such as laughter, pauses, and choking up. The company says it is trained entirely on human speech data, using only a single rack of NVIDIA GPUs for training, and can run locally on MacBooks with M4 Max and higher configurations. At this stage Dessa is not open for public trial; Block is sending testing invitations to security researchers, focusing on risk scenarios in which speech agents deceive humans.

Block releases full-duplex voice model Dessa, which can run locally on consumer-grade GPUs

Block's research lab has released the voice model Dessa, which uses a full-duplex architecture, can listen and speak at the same time, and can run locally on a MacBook M4 Max or a more powerful consumer-grade computer. The model was trained entirely on human speech and a single NVIDIA GPU, and the company says expressions such as emotion, pauses, and laughter are built into the model rather than driven by a text pipeline. Dessa is not currently publicly available, and Block is recruiting security researchers to test the risk that it could be used to deceive conversation partners.

Block releases speech model Dessa, trained only on human conversation and able to run locally

Block releases speech model Dessa, saying it is trained only on human conversation, without using text or pretrained models. Dessa uses a full-duplex approach, allowing it to keep listening while speaking and to interject, laugh, or respond; Block says it can run locally on MacBook Pro models with M4 Max and above chips. Dessa is currently a research preview, with no release date or pricing, and early access is open only to a small number of AI safety researchers.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Mistral's Voxtral 3 launches on OpenRouter, including two speech transcription models and one speech synthesis model

OpenRouter announced that Mistral's Voxtral 3 is now available. The series includes two new speech transcription models and one new speech synthesis model, covering more than 50 languages. Mistral officially states that Voxtral 3's Transcribe and Speech models power its new voice agents API.

English translation is being prepared.

Mistral Releases Voxtral 3 Speech Model with Real-Time Speaker Diarization and Turn Detection

Mistral has released Voxtral 3, comprising the Transcribe and Speech models, to power its new voice agents API. Co-founder Arthur Mensch added that Voxtral can perform speaker diarization and end-of-turn detection on the fly, and can be deployed either on your own infrastructure or via Mistral's cloud service.

Mistral releases the Voxtral 3 speech model series, including TTS and three transcription models

Mistral releases the Voxtral 3 speech model series, including one text-to-speech model and three speech-to-text models (Small Transcribe 3, Mini Transcribe 3, Mini Transcribe Realtime 3). Artificial Analysis evaluations show that Voxtral Small Transcribe 3 ranks 3rd on the speech-to-text leaderboard with a 2.1% AA-WER Index, ahead of ElevenLabs Scribe v2 at 2.2% and Grok Voice Transcribe 2.0 at 2.3%; Voxtral 3 TTS ranks 12th (out of 98) on the Provider Voice leaderboard with an Elo of 1,218, priced at $13/1M characters.