Search “Speech”

Matt Turck says open-source voice AI welcomes a small on-device SOTA TTS model suite

Matt Turck reposted and commented that a small on-device SOTA text-to-speech model suite has emerged in the open-source voice AI space. According to related reports, Gradium has open-sourced the Phonon on-device speech synthesis model suite, with all model parameters under 50m, focusing on multilingual naturalness and accuracy, and claiming to offer the broadest inference support to date; among them, Phonon-31m is released under the Apache 2.0 license. Currently, this news has been reposted by Matt Turck, with no further developments yet.

Matt Turck says open-source voice AI welcomes a small on-device SOTA TTS model suite

Matt Turck reposted and commented that a small on-device SOTA text-to-speech model suite has emerged in the open-source voice AI space. According to related reports, Gradium has open-sourced the Phonon on-device speech synthesis model suite, with all model parameters under 50m, focusing on multilingual naturalness and accuracy, and claiming to offer the broadest inference support to date; among them, Phonon-31m is released under the Apache 2.0 license. Currently, this news has been reposted by Matt Turck, with no further developments yet.

View story:Matt Turck says open-source voice AI welcomes a small on-device SOTA TTS model suite

HeyGen Professional Voice Clone is trained on 30 minutes to 2 hours of recorded audio

HeyGen has launched Professional Voice Clone, which is trained using 30 minutes to 2 hours of recorded audio. Related reports point out that this amount of data is key: short reference clips cannot capture the breathing rhythm before a long sentence or the accent features when speech speeds up, and these subtle irregularities are exactly what listeners use to judge whether a voice belongs to a real person. The reports also say that the clone preserves breathing, pauses, laughter, accent, and vowel pronunciation; HeyGen states that no one can tell the clone's result apart from the person themselves. Currently, this fact is supported only by this one report, with no subsequent developments.

Mistral releases the Voxtral 3 speech model series

On October 21, 2026, Mistral released the Voxtral 3 speech model series, including one text-to-speech model and three speech-to-text models: Small Transcribe 3, Mini Transcribe 3, and Mini Transcribe Realtime 3. Artificial Analysis's evaluation shows 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 Mini Transcribe 3 ranks 5th at 2.2%. Voxtral 3 TTS has an Elo of 1,218 in the provider voice arena, ranking 12th among 98 models, and is priced at $13 per million characters. There is currently no further evidence-based development.

Vapi Hosts Fleet Week Voice AI Event in San Francisco Bay

On October 8, Vapi chartered a yacht in San Francisco Bay to host a Fleet Week voice AI event, with more than 30 voice AI builders in attendance; Deepgram, Kouper Health, and Coval hosted fireside chats and technical workshops. According to the Vapi Blog, the most-discussed topic on board was GPT Live: skipping traditional transcription/synthesis can reduce latency and make conversational turn-taking smoother, but it comes at the cost of fine-grained control over key terms and scripts. The focus of discussion has shifted from "whether voice AI can hold a conversation" to "how to make conversations more natural," specifically including backchanneling, perceived latency, and measuring naturalness. Currently, this event is disclosed by a single source, the official Vapi blog, and there are no subsequent factual developments.

Scaled Cognition launches voice model Voice

On October 20, 2026, Scaled Cognition Inc. launched its text-to-speech model Voice. According to a SiliconANGLE report the same day, the model supports more than 25 languages and voice cloning, can be deployed via API calls or self-hosted, and focuses on accurate reading of details such as account numbers and addresses. The report said hosted API pricing is $9 per million characters, while self-hosting is $1 per million characters (excluding infrastructure); the company says that at a scale of 5 million characters, including infrastructure, the cost is about $13 per million characters, and it claims a speech generation latency of 70 milliseconds. Its Smart Pronunciation capability first determines the meaning of abbreviations and numbers before generating audio—for example, reading "Pay 1/2 by 1/2" as "pay half by January second"—and it can run without relying on APT or other large language models. There are currently no subsequent developments.

inclusionAI open-sources the Ming-3.0-asr-tiny speech recognition model repository

inclusionAI has open-sourced the Ming-3.0-asr-tiny speech recognition model repository on GitHub. The model combines the Ling-tiny-v3 language model with an audio encoder, supporting Chinese-English transcription, dialect recognition, and Context ASR. Dialect evaluation covers 15 test sets, including Hunan, Minnan, Cantonese, and Sichuan-Chongqing; Context ASR can recognize proper nouns through hotwords in prompts or prior conversation. The inference stack uses Transformers 5.12.1, fla-core 0.5.2, and PyTorch 2.11.0, and provides an SGLang accelerated deployment example. There are currently no subsequent development facts.

inclusionAI open-sources the Ming-3.0-asr-tiny speech recognition model repository

inclusionAI has open-sourced the Ming-3.0-asr-tiny speech recognition model repository on GitHub. The model combines the Ling-tiny-v3 language model with an audio encoder, supporting Chinese-English transcription, dialect recognition, and Context ASR. Dialect evaluation covers 15 test sets, including Hunan, Minnan, Cantonese, and Sichuan-Chongqing; Context ASR can recognize proper nouns through hotwords in prompts or prior conversation. The inference stack uses Transformers 5.12.1, fla-core 0.5.2, and PyTorch 2.11.0, and provides an SGLang accelerated deployment example. There are currently no subsequent development facts.

View story:inclusionAI open-sources the Ming-3.0-asr-tiny speech recognition model repository

Kotoba Speech API Now Live on fal

fal announced that Kotoba's speech API is now live. According to information released by fal, the API supports real-time speech translation (text and audio output), streaming audio transcription, speech generation with preset voices, and recording transcription with optional word-level timestamps. Currently, this event has only been reported by fal as a single source, and there are no subsequent developments or facts.

HeyGen Professional Voice Clone is trained on 30 minutes to 2 hours of recordings

HeyGen has launched Professional Voice Clone, which is trained using 30 minutes to 2 hours of recorded audio. Related reports point out that this amount of data is key: short reference clips cannot capture the breathing rhythm before a long sentence or the accent features when speech speeds up, and these subtle irregularities are exactly what listeners use to judge whether a voice belongs to a real person. The reports also say that the clone preserves breathing, pauses, laughter, accent, and vowel pronunciation; HeyGen states that no one can tell the clone's result apart from the person themselves. Currently, this fact is supported only by this one report, with no subsequent developments.

View story:HeyGen Professional Voice Clone is trained on 30 minutes to 2 hours of recorded audio

Mistral releases the Voxtral 3 speech model series

On October 21, 2026, Mistral released the Voxtral 3 speech model series, including one text-to-speech model and three speech-to-text models: Small Transcribe 3, Mini Transcribe 3, and Mini Transcribe Realtime 3. Artificial Analysis's evaluation shows 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 Mini Transcribe 3 ranks 5th at 2.2%. Voxtral 3 TTS has an Elo of 1,218 in the provider voice arena, ranking 12th among 98 models, and is priced at $13 per million characters. There is currently no further evidence-based development.

View story:Mistral releases the Voxtral 3 speech model series

Ant Ling Open-Sources the Ming-3.0 Speech Series, Five Models in Total

On October 20, 2026, Ant Ling (inclusionAI) open-sourced the Ming-3.0 speech series on GitHub, offering three series—speech synthesis, audio understanding and dialogue, and speech recognition—for a total of five models. Among them, Ming-3.0-tts includes a 0.5B dense version and a tiny MoE version based on Ling V3 with 7.9B total parameters and 1.3B activated parameters, supporting voice cloning, text-designed timbres, and multilingual synthesis; Ming-3.0-audio-flash uses a 126B-parameter MoE architecture with 6.5B activated parameters, and there is also a tiny version with 9.3B parameters and 2.7B activated parameters, covering multi-dialect accent recognition, long audio understanding, and end-to-end speech dialogue, under the MIT license. On the same day, vLLM announced that vLLM-Omni would provide support on the day of release, and Ant Ling thanked it and said deployment would be more convenient; atomic.chat will also integrate Ming-3.0 into local AI workflows. Artificial Analysis published an evaluation stating that Ming-3.0-tts ranks first among open-weight models in the controlled speech TTS Arena, and its pronunciation robustness of 93.2% ties for first among 31 models. On October 22, a blogger said they planned to trial-run the 0.5B version locally; on October 23, Artificial Analysis said it achieved top-tier TTS quality with 0.5B parameters.

English translation is being prepared.

Vapi hosts Fleet Week voice AI event in San Francisco Bay

On October 8, Vapi chartered a yacht in San Francisco Bay to host a Fleet Week voice AI event, with more than 30 voice AI builders in attendance; Deepgram, Kouper Health, and Coval hosted fireside chats and technical workshops. According to the Vapi Blog, the most-discussed topic on board was GPT Live: skipping traditional transcription/synthesis can reduce latency and make conversational turn-taking smoother, but it comes at the cost of fine-grained control over key terms and scripts. The focus of discussion has shifted from "whether voice AI can hold a conversation" to "how to make conversations more natural," specifically including backchanneling, perceived latency, and measuring naturalness. Currently, this event is disclosed by a single source, the official Vapi blog, and there are no subsequent factual developments.

View story:Vapi Hosts Fleet Week Voice AI Event in San Francisco Bay

Scaled Cognition launches voice model Voice

On October 20, 2026, Scaled Cognition Inc. launched its text-to-speech model Voice. According to a SiliconANGLE report the same day, the model supports more than 25 languages and voice cloning, can be deployed via API calls or self-hosted, and focuses on accurate reading of details such as account numbers and addresses. The report said hosted API pricing is $9 per million characters, while self-hosting is $1 per million characters (excluding infrastructure); the company says that at a scale of 5 million characters, including infrastructure, the cost is about $13 per million characters, and it claims a speech generation latency of 70 milliseconds. Its Smart Pronunciation capability first determines the meaning of abbreviations and numbers before generating audio—for example, reading "Pay 1/2 by 1/2" as "pay half by January second"—and it can run without relying on APT or other large language models. There are currently no subsequent developments.

View story:Scaled Cognition launches voice model Voice

Ant Ling Open-Sources the Ming-3.0 Speech Series Models

On October 20, 2026, Ant Ling (inclusionAI) open-sourced the Ming-3.0 speech series on GitHub, offering three series—speech synthesis, audio understanding and dialogue, and speech recognition—for a total of five models. Among them, Ming-3.0-tts includes a 0.5B dense version and a tiny MoE version based on Ling V3 with 7.9B total parameters and 1.3B activated parameters, supporting voice cloning, text-designed timbres, and multilingual synthesis; Ming-3.0-audio-flash uses a 126B-parameter MoE architecture with 6.5B activated parameters, and there is also a tiny version with 9.3B parameters and 2.7B activated parameters, covering multi-dialect accent recognition, long audio understanding, and end-to-end speech dialogue, under the MIT license. On the same day, vLLM announced that vLLM-Omni would provide support on the day of release, and Ant Ling thanked it and said deployment would be more convenient; atomic.chat will also integrate Ming-3.0 into local AI workflows. Artificial Analysis published an evaluation stating that Ming-3.0-tts ranks first among open-weight models in the controlled speech TTS Arena, and its pronunciation robustness of 93.2% ties for first among 31 models. On October 22, a blogger said they planned to trial-run the 0.5B version locally; on October 23, Artificial Analysis said it achieved top-tier TTS quality with 0.5B parameters.

View story:Ant Ling Open-Sources the Ming-3.0 Speech Series, Five Models in Total

Artificial Analysis announces Ming-3.0-tts speech evaluation results

On October 20, 2026, Ant Ling (inclusionAI) open-sourced the Ming-3.0 speech series on GitHub, offering three series—speech synthesis, audio understanding and dialogue, and speech recognition—for a total of five models. Among them, Ming-3.0-tts includes a 0.5B dense version and a tiny MoE version based on Ling V3 with 7.9B total parameters and 1.3B activated parameters, supporting voice cloning, text-designed timbres, and multilingual synthesis; Ming-3.0-audio-flash uses a 126B-parameter MoE architecture with 6.5B activated parameters, and there is also a tiny version with 9.3B parameters and 2.7B activated parameters, covering multi-dialect accent recognition, long audio understanding, and end-to-end speech dialogue, under the MIT license. On the same day, vLLM announced that vLLM-Omni would provide support on the day of release, and Ant Ling thanked it and said deployment would be more convenient; atomic.chat will also integrate Ming-3.0 into local AI workflows. Artificial Analysis published an evaluation stating that Ming-3.0-tts ranks first among open-weight models in the controlled speech TTS Arena, and its pronunciation robustness of 93.2% ties for first among 31 models. On October 22, a blogger said they planned to trial-run the 0.5B version locally; on October 23, Artificial Analysis said it achieved top-tier TTS quality with 0.5B parameters.

View story:Ant Ling Open-Sources the Ming-3.0 Speech Series, Five Models in Total

English translation is being prepared.

View story:English translation is being prepared.

ElevenLabs and KITS sign memorandum of cooperation on Kannada speech AI

On October 19, 2026, ElevenLabs and the Karnataka Innovation and Technology Society (KITS) signed a memorandum of understanding to explore the use of Kannada speech AI for citizen services, skills training, and grievance handling in Karnataka, and to enable users who have lost their voice to communicate via synthesized speech and receive free lifetime licenses. Reports from representatives said that Indian speech AI startups opposed this: Dheemanth Reddy of Maya Research is concerned that sensitive conversations between citizens and the government will flow to foreign jurisdictions, and Raoul Nanavati of Navana.ai believes that government contracts are the most valuable endorsement for local companies. The same report also compared prices: the discounted price of the ElevenLabs v4 text-to-speech API is about 2.11 rupees per 1,000 characters, while Sarvam AI is 3 rupees per 1,000 characters. There are currently no subsequent developments of fact.

ElevenLabs and KITS sign memorandum of cooperation on Kannada speech AI

On October 19, 2026, ElevenLabs and the Karnataka Innovation and Technology Society (KITS) signed a memorandum of understanding to explore the use of Kannada speech AI for citizen services, skills training, and grievance handling in Karnataka, and to enable users who have lost their voice to communicate via synthesized speech and receive free lifetime licenses. Reports from representatives said that Indian speech AI startups opposed this: Dheemanth Reddy of Maya Research is concerned that sensitive conversations between citizens and the government will flow to foreign jurisdictions, and Raoul Nanavati of Navana.ai believes that government contracts are the most valuable endorsement for local companies. The same report also compared prices: the discounted price of the ElevenLabs v4 text-to-speech API is about 2.11 rupees per 1,000 characters, while Sarvam AI is 3 rupees per 1,000 characters. There are currently no subsequent developments of fact.

View story:ElevenLabs and KITS sign memorandum of cooperation on Kannada speech AI