Microsoft · MIT

VibeVoice-ASR

Microsoft’s 8.7B model that transcribes up to an hour in one pass, with who-said-what and timings.

How much memory it needs

Memory to run each model, from its model card or the size of its files. Smaller versions trade some quality for less memory.

ModelVersionNeedsBusiness use
VibeVoice-ASRBitNet (CPU)1.6 GBAllowed
BF1617.3 GB

Which computers run it

The best model in this group for each machine.

ComputerBest versionFit
Laptop or PC, no graphics card, 32 GBVibeVoice-ASR BF16Runs well
RTX 4060 · 8 GBVibeVoice-ASR BitNet (CPU)Runs great
RTX 3060 · 12 GBVibeVoice-ASR BitNet (CPU)Runs great
RTX 5060 Ti · 16 GBVibeVoice-ASR BitNet (CPU)Runs great
RTX 4090 · 24 GBVibeVoice-ASR BF16Runs well
RTX 5090 · 32 GBVibeVoice-ASR BF16Runs great
Mac with M4 · 16 GBVibeVoice-ASR BitNet (CPU)Runs great
Mac with M4 Pro · 48 GBVibeVoice-ASR BF16Runs great
Mac with M5 Max · 128 GBVibeVoice-ASR BF16Runs great
Ryzen AI Max+ 395 · 128 GBVibeVoice-ASR BF16Runs great
DGX Spark · 128 GBVibeVoice-ASR BF16Runs great
RTX PRO 6000 · 96 GBVibeVoice-ASR BF16Runs great
Mac Studio with M5 Ultra · 256 GBVibeVoice-ASR BF16Runs great
Mac Studio with M3 Ultra · 512 GBVibeVoice-ASR BF16Runs great

What it is good and bad at

Good at

  • Speaker labels and timestamps
  • 50+ languages

Not good at

  • About 17 GB at full quality

Start it

  • VibeVoice-ASR: runs in Transformers (also VibeASR.cpp (BitNet, CPU)). Model card