Unverified50% confidenceBenchmarkExact time
The paper separates the decoder gain from the transposed-direction read gain: with forward reads only, it uses 1.6 to 4.1 times fewer reads than the hybrid baseline; combined decoding plus transposed reads uses a further 1.3 to 4 times fewer reads at 1% noise, and the second-direction gain shrinks as noise rises.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/21/2026
First Seen
Valid until: 1/19/2027
Sources
Related Entities
Related Claims
UnverifiedText-based methods flatten the word frequency distribution, increasing the occurrence rate of low-frequency features by 37% to 38%, thereby reversing errors on tones and palatalization.74% similarUnverifiedIn speech recognition, encoder weights are generally more sensitive to quantization, while the decoder is relatively more tolerant of precision loss.69% similarUnverifiedNon-streaming encoders can use bidirectional attention or global convolution, generating representations after seeing the complete speech sequence, and typically achieve higher accuracy.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1080728API
curl https://kongchang.com/api/v1/knowledge/claims/1080728MCP
get_claim(id=1080728)