Microsoft ships streaming speech models as AI video hype meets a sceptical crowd

Microsoft AI released MAI-Transcribe-2-Streaming and two text-to-speech models on 2 October, according to The Decoder, which reviewed the launch. Microsoft says the transcription model covers 60 languages and returns partial results in just over 100 milliseconds.

That is not a video generator. But it is the layer that decides whether a synthetic presenter feels like a conversation or a recording, and it lands in the same week that the video side of the industry got a lesson in how hard the last mile is.

The realism claim that is easiest to check

Tavus introduced Griffin on 1 October, calling it the first "Human Interaction Model". The Decoder reported the company's own study: 48 percent of participants believed Griffin was a real person after a one-minute video call, against a previous best of 2 percent. Tavus says an Nvidia test scored Griffin 3.83 points for how human an AI feels in audio-video conversation, with actual humans at 3.92 and the previous best AI model at 2.80.

Only a preview, Griffin-Lite, is available to selected testers, with a fuller version promised once safety concerns are addressed. Tavus was founded in 2020 and has raised about $64 million.

Compare that with Microsoft's voice numbers. MAI-Transcribe-2-Streaming costs $0.54 per hour of audio at an introductory price through the end of the year. MAI-Voice-2.1 speaks 23 languages in the same voice with a native accent in each, according to Microsoft, and the Flash variant costs $15 per million characters instead of $22. Both can clone a voice from a few seconds of reference audio. In one Microsoft test, about half of 4,000 participants thought the voices belonged to a real person.

Video models keep shipping, and so do the questions

The wider release calendar is crowded. The dossier's headline list includes FLUX 3 Image, described as supporting up to 4K generation, subject positioning and partial editing, with an open-source version expected in weeks, plus ByteDance's Seedance 2.5 and Kling 4.0 from Kuaishou's AI spinoff. None of those releases is detailed in the source texts supplied here, so the specifics are thin. What is documented is the adjacent tooling.

Ideogram said on 1 October that Ideogram 4.5 will only touch the part of an image a user selects, at four quality tiers from 0.8 to 22 cents per image and native 2K resolution, with an open-weight release promised. That is a correction to a known failure mode rather than a new capability: repeated edits on models such as GPT-Image 2.5 and Nano Banana still leave artifacts, per The Decoder.

Then there is the money. Dealroom reported on 30 September that Luma scaled 9,000 GPUs in six hours for a Dream Machine launch. Runway moved into robotics with an open-weight model called Praxis-1, according to Robotics & Automation News on 1 October. Both sit outside the ten-source core below because the dossier carries only their headlines.

"Storage systems smooth out prices, but they don't generate electricity," Leonhard Gandhi, project manager for Energy Charts at Fraunhofer ISE, said as his institute published a simulator for German day-ahead power prices.

He was talking about batteries, not GPUs. The analogy is not perfect, but the constraint is the same shape: inference capacity is a physical input, and the dossier shows what happens when it becomes scarce. Denmark's parliament adopted an emergency grid plan on 2 October that replaces first-come, first-served connections with four priority categories, and data centers sit in the least prioritised tier unless tied to a critical societal function.

Meanwhile the model layer keeps fragmenting. Amazon Web Services released Strands Decider 2B, an open-source decision model, in the same week OpenAI announced a similar offering, TechCrunch reported on 1 October. Cloudflare answered with Clef and Clef-flash, open-weight decision models it says beat TypeSafe's Jev on three of four of TypeSafe's own benchmarks, at $0.24 per million tokens against Jev's $0.042. Cloudflare self-reported those scores and they have not been reproduced on the official index, as The Register noted.

None of it settles the question a viewer asks in the first two seconds of a synthetic clip: is this a person. Tavus has one number for that. It is 48 percent, on a one-minute call, in a study the company ran itself.