Microsoft opens VibeVoice for 60-minute ASR and 300ms streaming TTS

• Microsoft opened VibeVoice as an open-source voice stack with long-form ASR and streaming TTS; the launch thread says it can transcribe 60-minute recordings in one pass and generate first audio in about 300ms.
• The same repo link post points to structured transcripts with speaker labels and word-level timestamps, which makes the release more useful for podcast edits, interview logging, and dialogue-heavy production than plain text dumps.
• VibeVoice also targets multilingual workflows: the launch thread says it supports 50-plus languages, and the [img:0|VibeVoice graphic] shows Microsoft positioning ASR and TTS as one model family.
• A parallel HN thread around Kitten TTS shows where creator interest is heading: smaller, CPU-friendly voice models are attractive for narration and bots, but the HN discussion says latency, streaming architecture, and dependency bloat still decide whether a voice stack is usable on edge hardware.

Microsoft's VibeVoice release is two products in one family: long-context speech recognition and low-latency text-to-speech. In the launch thread, Hasan Toor summarizes the headline numbers as 60-minute single-pass ASR, speaker identification, word timestamps, and roughly 300ms first-audio latency for streaming TTS. The linked GitHub repo adds that the ASR side can output structured transcriptions with timestamps and speaker attribution, and that the stack supports fine-tuning, vLLM-based inference, and 50-plus languages.

For creators, the practical shift is less cleanup between recording and edit. A single recording can be turned into a speaker-separated transcript for rough cuts, while the TTS side is aimed at fast voice generation instead of offline batch export. The [img:0|project graphic] also shows Microsoft framing VibeVoice as a broader open-source voice platform rather than a single demo model.

VibeVoice lands into a voice market that is splitting in two directions: bigger open models that cover more of the workflow, and tiny models that can run closer to the device. The Kitten TTS page describes Kitten TTS as an ONNX library with 15M to 80M parameter models, CPU inference, eight built-in voices, adjustable speed, and 24 kHz output, with the smallest model under 25MB.

That contrast sharpens the real question for creatives: not just output quality, but where the model can run and how fast it starts talking. In the HN thread, one user reports using Kitten in a Discord bot workflow at about 1.5x realtime on an Intel 9700 CPU, while the HN discussion highlights the recurring blockers on lightweight stacks: installation pulling heavy dependencies, unclear first-chunk latency, and streaming limitations on low-power hardware.