MICROSOFT / LONG-FORM ASR / 2026

VibeVoice-ASR

TOTAL8.33Bparameters

ACTIVE8.33Bruntime path

DEPTH2×7 + 28repeated blocks

HIDDEN3,584main trunk

CONTEXT131Kstream / sequence

CHECKPOINT16.67GB8 shards · MIT

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

3,600 seconds → 4,096 tokens · 854.59T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT3.6KSECONDS

TRANSFORMER STACKDUAL ENCODER + 28Laggressive audio downsampling → long-context decoder

Pre-Norm Residual

OUTPUT4.1KTOKENS

wraps every repeated block

LONG-CONTEXT GQA · SINGLE-PASS HOUR通过激进声学降采样与 131K 文本上下文，在一次前向中做长音频转写。

Diarization Head解码器把说话人、时间和文本统一生成为结构 token。

AUDIO60minsingle pass

DECODER28L3,584

HEADS28 / 4Q / KV

CONTEXT131Ktokens

VOCAB152Kstructured

PARAMS8.33Bheader