• Streaming sherpa-onnx models no longer cut off the last word or two of an
  utterance. The trailing silence fed before input_finished() is now measured
  from the model rather than fixed at the 0.66s in sherpa-onnx's own examples:
  the encoder only runs on whole chunks, and a chunk is 1.41s for the Kroko
  streaming zipformers that --sherpa-streaming defaults to, so the tail of
  every utterance went undecoded. 0.66s stays the minimum for models whose
  chunk is shorter
  • Priming those models at startup was a no-op for the same reason — it fed
    128 samples, far less than one chunk — so the first transcription paid for
    the warm-up