Audio & Speech

Today's Stories

1. 1

   Voice AI brings hands-free ops to central kitchens

   Advanced Media and AWS Japan co-wrote this post about a solution shown at AWS Summit Japan 2026. It pairs the AmiVoice speech recognition engine with AI agents built on Amazon Bedrock AgentCore Runtime and the Strands Agents SDK. A serverless backend using Amazon DynamoDB and AWS Lambda completes the architecture for hands-free recipe operation. Central kitchen workers often have both hands occupied and must remove sanitary gloves to check recipes or log inventory, causing interruptions and hygiene risk. Manual record-keeping also leaves digital data uncollected, which limits management visibility; the solution records structured data automatically in the background as staff speak.

   AmiVoice also offers on-device and on-premises SDK options for environments where audio cannot leave the site, although AWS Marketplace currently lists only the cloud AmiVoice API. The post frames the pattern as applicable beyond food, with AI agents expected to perform tasks autonomously in next-generation workflows.

2. 2

   Meta Releases Muse Voice Transcribe, a Real-Time Speech Recognition Model

   Meta has announced Muse Voice Transcribe, its first real-time speech recognition model developed by Meta Superintelligence Labs (MSL). It is available from September 1 through the Meta Model API at $0.18 per hour. The model can distinguish more than 20 speakers and handle audio over an hour without post-processing, using a single model for streaming recognition, speaker separation, and end-pointing. It was trained on 70+ languages, with 25 verified at launch, including Japanese, and natively handles code-switching.

   Meta claims it ranks #1 in Artificial Analysis' streaming speech recognition and public speaker separation benchmarks. Its word error rate is 3.1% for finalized transcripts, with a 0.16-second delay from end of speech, beating Cartesia Ink-2 (3.4%) and ElevenLabs' Scribe v2 Realtime (3.6%). The model is also integrated into Mac's Meta AI and Muse Code for voice input.

3. 3

   Winamp Group's Jamendo expands AI music lawsuits, adds six more targets

   Winamp Group's subsidiary Jamendo SA amended its U.S. federal complaint against NVIDIA, which has now been formally served; the first hearing is scheduled for January 12, 2027. Separately, Jamendo withdrew its complaint against Suno and plans to refile an updated, more comprehensive version shortly. Jamendo is broadening its fight over alleged unauthorized use of musical content and data in AI development. Beyond NVIDIA and Suno, it has identified six additional technology and AI companies that it says used its content or data, and has already contacted some of them.

   Jamendo says its priority is a commercial resolution — compensation for past uses plus licensing or partnerships for future use — and it intends to enforce its rights where talks fail, as it has in the NVIDIA and Suno cases.

4. 4

   Google、Gemini 3.7 Flash公開、Pixel 11発表

   Google released Gemini 3.7 Flash, its latest developer model, and unveiled the new Pixel 11 series. The Gemini app officially surpassed 1 billion monthly users. Gemini 3.7 Flash arrived just three weeks after 3.6 Flash at half the introductory price per million tokens. The Pixel 11 devices run the Google Tensor G6 chip with the latest Gemini Nano model.

   Gemini Omni 1.1 Flash now supports scene extension, 4K upscaling, and faster prototyping. Gemini 3.5 Transcribe offers real-time transcription for voice agents and live captioning.

5. 5

   Phonely launches Alma, voice AI trained on 10M calls

   Phonely Ltd. launched Alma, a large language AI model built for voice agents and trained on over 10 million real phone conversations. Alma provides sub-185-millisecond response times to the first token, compared with roughly 500 milliseconds for OpenAI's GPT-4.1. Many AI models for voice agents are trained on text, which can make conversations sound unnatural. Alma reduces pauses that interrupt conversations and is pre-trained to sound like a person, cutting the need for prompt engineering. The model also costs 55 cents per blended million tokens, making it 84% cheaper than GPT-4.1 and 90% cheaper than GPT-5.4.

   Alma uses real traffic to improve how it interacts with calls without waiting for the next training cycle, aiming to deliver better results the next day. It can work with existing transcribers and text-to-speech providers, and is already answering millions of Phonely calls monthly.

6. 6

   TontaubeV1 released: character-level TTS for long-form

   Two developers released TontaubeV1, a 2.9B-parameter open-weight TTS model for expressive speech and long-form narration. It supports zero-shot voice cloning from up to one minute of reference audio and targets English and German. The model uses character-level tokenization, which the authors found works better than the original BPE tokenizer in early experiments. This choice, plus building on the DualCodec multi-codebook audio codec, aims to improve long-form generation and low-latency local inference.

   The model was trained on 7 languages and about 200k hours of audio, but is mostly tested in English and German. Availability of the open weights and the specific licensing or access details were not mentioned in the announcement.