What happened
The encoder produced 6.5 audio tokens per second, while the pipeline expected 6.25. The team retrained 32B-Para v1.5 and 4B-Para v1.4 with the full input. In its English test at a 300 ms pause, premature cutoffs fell from 20.4% to 17.7% for the larger model and from 25.8% to 21.8% for the smaller one.
Why it matters
These are developer measurements, not interruption rates on real calls. The Para weights carry a noncommercial CC BY-NC 4.0 license; the Apache license for the code does not grant commercial use of those weights.
What to watch
Test pauses, hesitation and background noise on representative calls before treating the benchmark improvement as a deployment result.
Sources & verification
Source links are not a count of independent confirmations.
01Report a correction
Help us get it right
Include the disputed statement and, if possible, a supporting source URL. Please do not send personal or sensitive information.
