Nvidia released Nemotron 3.5 Transcribe, a free, open-weight transcription model supporting up to 40 languages, alongside version 2 of its open-source Granary speech dataset. The company says the 30-billion-parameter model can translate 31 languages to and from English, process recordings up to an hour long, distinguish up to eight speakers, and run on a single chip ranging from Jetson AGX Thor to Blackwell datacenter hardware. Nvidia claims it outperforms rivals on benchmarks including Hugging Face's Open ASR Leaderboard, though it only works on recorded audio rather than live streams.
techcrunch.com
· 2026-10-28
BetterWispr is a new Apache 2.0-licensed dictation tool for macOS 14 and later that requires no subscription, account or API key. It runs speech-to-text locally using downloadable models or Apple's on-device speech recognition, letting users dictate into any text field via a hold-to-record shortcut like Option+Space.
betterwispr.com
· 2026-10-11
A new speech recognition model called Whistle has been released for mobiles, wearables, robots, smart home devices, automotive systems and microcontrollers. It ships as a single 16.9 MB file that runs entirely on-device via CPU with no dependencies, handling transcription, word-level timestamps and speech embeddings for seven languages without sending audio off the device.
cactuscompute.com
· 2026-10-08
Translation earbuds, such as Apple's AirPods Pro paired with the Translate app, let users hear real-time spoken translations by connecting to a phone, selecting languages and activating a live translation mode. Some designs require sharing an earbud with the other speaker instead. Behind this simple interface, the devices run three technical stages: speech recognition, machine translation, and text-to-speech synthesis.
engadget.com
· 2026-09-29
Wispr Advanced Interfaces Lab has released Canto, a new speech recognition model built specifically for real-time dictation in noisy, everyday conditions rather than clean studio audio. In tests using 10 hours of real Wispr Flow dictations from over 2,300 speakers, Canto posted the lowest word error rate compared with models from Google, OpenAI, AssemblyAI, and Deepgram.
wisprflow.ai
· 2026-09-17
Microsoft AI launched MAI-Transcribe-2, a speech-to-text model it says beats OpenAI, Google and ElevenLabs on speed, accuracy and cost, pricing it at 10 cents per audio hour—a 72% cut from the $0.36 rate it charged for its first model just five months ago. The new version supports 60 languages, up from 43, and adds built-in speaker diarization and word-level timestamps, targeting noisy real-world business audio rather than clean recordings.
venturebeat.com
· 2026-09-03
Google unveiled Gemini 3.5 Transcribe, a new speech-to-text AI model that detects over 85 languages, strips filler words, and reformats casual speech into clean, structured text. It can also learn custom vocabulary, capture alphanumeric strings like order numbers, and assign dialogue to up to three speakers with timestamps from recorded audio.
engadget.com
· 2026-08-26
Google has released Gemini Audio, a set of three new speech-focused AI models: Gemini 3.5 Live, 3.5 Live Experimental, and 3.5 Transcribe. These models are meant to handle real-time conversation and speech-to-text tasks and are rolling out across Search Live, Gemini Live, Docs, Keep, Gmail, the Gemini app, and Gboard. Notably, Gemini 3.5 Transcribe replaces Google's older Chirp 3 model and adds features like filler-word removal, auto-formatting, and voice editing.
androidauthority.com
· 2026-08-26
GoKawiil