Deep Learning Weekly: Issue 449
Deep Learning Weekly Miko Planas ● Covered by 2 sources
This week's roundup: Google, Mistral, and Cohere all dropped new audio AI models, plus a clever paper on speeding up sparse attention. Voice AI is quietly becoming the next battleground while researchers chip away at making inference cheaper.
Based on reporting by Deep Learning Weekly, Miko Planas — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Audio AI had a genuinely busy week, and it's worth pausing on why three separate labs pushed out speech models within days of each other. Google shipped Gemini 3.1 Flash Live, its most capable real-time audio model yet, hitting 90.8% on ComplexFuncBench Audio and 36.1% on AudioMultiChallenge — numbers that matter if you're building voice agents that need to hold a conversation without falling apart mid-sentence. Mistral countered with Voxtral TTS, a compact 4-billion-parameter text-to-speech model that covers nine languages, clones a voice from just three seconds of audio, and responds in roughly 70 milliseconds. Cohere, meanwhue, went after transcription rather than generation: its new Transcribe model tops the HuggingFace Open ASR Leaderboard with a 5.42% average word error rate across 14 languages, and it's fully open-source at 2 billion parameters.
That last detail is the interesting part. A model small enough to run without a data center's worth of GPUs is now state-of-the-art on a public benchmark, which says something about how fast the efficiency gap is closing between open and closed audio systems. Meta didn't sit this one out either, updating its segmentation tool SAM 3.1 with object multiplexing that doubles video processing speed to 32 frames per second on a single H100. Video tracking, transcription, synthesis — the entire audio-visual stack got faster this week, not just smarter.
On the money side, Granola raised a $125 million Series C at a $1.5 billion valuation, led by Index Ventures, after posting 250% revenue growth in a single quarter. The company started as an AI note-taker for meetings and is now talking about pushing into agentic task automation, which is the same trajectory nearly every productivity AI startup seems to be on right now: start narrow, then try to become the assistant that runs your whole workday.
And then there's the research end of things, which is less flashy but arguably more consequential long-term. A new paper called IndexCache tackles a real bottleneck in sparse attention systems like DeepSeek's DSA, where a lightweight indexer has to run at every single layer even though the top-k token selections barely change from one layer to the next. By letting most layers simply reuse the nearest full layer's index rather than recomputing it, the authors cut 75% of indexer computations on a 30-billion-parameter model with almost no quality loss, delivering up to 1.82x faster prefill and 1.48x faster decoding. Early tests on GLM-5 suggest the trick scales too, which is the kind of unglamorous engineering that actually determines how much your inference bill looks like next quarter.
My take — AI-written commentary, not fact-checked reporting
The audio race this week tells you where the next fight is heading — not chatbots, but always-on voice agents that need to be fast, cheap, and multilingual, and open models like Cohere's Transcribe are closing the gap on closed ones faster than people expect. I'll also take IndexCache over another benchmark-topping foundation model any day; shaving 75% off indexer compute with no training required is the boring, unsexy work that actually keeps inference costs from eating every startup's margin alive.
Read more about this at: Deep Learning Weekly