FunASR’s new speech-recognition release keeps a result record for every recording even when its Paraformer model produces no text, preserving the link with timestamps. The update also respects a configured speaker limit and validates extra model settings in the realtime server. Developer checks use controlled examples and do not establish broader recognition-quality gains.
FunASR has released a correction to the way its speech-recognition toolkit returns a batch of results. In version 1.4.19, published on November 1, the Paraformer model keeps a keyed result for an utterance even when it predicts no text tokens. Retaining that entry preserves the correspondence between the input recording, its output and its timestamps. The release notes identify both empty-result handling and timestamp alignment as the scope of this repair.[1]
The update also makes automatic speaker clustering respect an explicit maximum speaker count. In the realtime server, additional Hugging Face model settings can be supplied as a JSON object, a structured format for exchanging data, and are validated before being forwarded. Separate ONNX source corrections are present in the Git tag, but the release does not publish a new funasr_onnx package. Python developers can obtain the packages from PyPI or GitHub.[1]
The developers describe checks of installed packages and controlled fixtures, rather than a general measurement of recognition accuracy on real recordings. They do not claim confirmation on the original recordings supplied by the people who reported the problems. The notes retain collection errors and skipped checks from an attempted full repository suite. Attached native runtime archives reuse an earlier release; those binaries were not rebuilt or executed for this publication.[1]
