Microsoft releases VibeVoice-ASR-Streaming: transcription emits speaker, time and content per chunk as speech arrives
Synopsis
Microsoft released VibeVoice-ASR-Streaming, which turns the "who said what" transcription that previously required processing a whole audio segment into continuous output as speech arrives. The earlier VibeVoice-ASR handled 60-minute long-form audio in a single pass, producing structured transcriptions with speaker, timestamps and content within a 64K token length. The new model emits text once per chunk and supports customized hotwords and 10 languages.
Interpretation
Microsoft changed VibeVoice-ASR transcription from a single pass over a whole audio segment to streaming, continuously emitting "who said what" while speech is still arriving. The earlier VibeVoice-ASR was designed for single-pass long-form processing, accepting up to 60 minutes of continuous audio within a 64K token length to keep speaker tracking and semantic consistency. The release states the contrast in processing timing: the earlier path transcribed after audio was assembled into a segment, while the new model produces text per chunk during arrival.
The streaming version keeps the unified speech-to-text model's structured output target, giving speaker, timestamps and content together. That output form comes from VibeVoice-ASR, whose structured transcription contains Who, When and What and supports user-customized context. The release says the new model supports customized hotwords and 10 languages, and provides demo and model entries.
Perspective
The demonstrated scope is Microsoft's released VibeVoice-ASR-Streaming model with its demo and model entries, supporting customized hotwords and 10 languages; it suits engineering readers interested in speech-transcription integration.
The release gives no transcription accuracy, latency or comparison against whole-segment processing in streaming mode, so actual behavior depends on the evaluations in the model and report.