FUNASR / NON-AUTOREGRESSIVE RICH ASR / 2024

SenseVoice-Small

TOTAL≈234Mparameters

ACTIVE≈234Mruntime path

DEPTH50 + 20repeated blocks

HIDDEN512main trunk

CONTEXTlong audiostream / sequence

CHECKPOINT≈930MB1 shards · FunASR Model License

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

30 seconds → 256 tokens · 600.00G arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT30SECONDS

TRANSFORMER STACK50L SAN-M + 20L TPencoder-only recognition with temporal predictor

Encoder Residual

OUTPUT256TOKENS

wraps every repeated block

SAN-M SELF-ATTENTION · RICH TRANSCRIPTION TOKENS同一次非自回归识别输出语言、情绪和环境事件标记。

Temporal Predictor不需要逐 token decoder，输出延迟主要由 encoder forward 决定。

ENCODER50LSAN-M

TP20Lpredictor

HIDDEN512output

HEADS4attention

FFN2,048dim

PARAMS≈234Msmall