toolmicrosoftvoicetranscriptionopen-sourceaudiospeech-to-texttext-to-speechasrttsUpdated 2026-07-01

VibeVoice

Microsoft's open-source voice AI family. The bookmark's durable claim maps to VibeVoice-ASR, which handles 60-minute long-form speech recognition in one pass with speaker, timestamp, and transcript structure.

View original tweet

What It Is

VibeVoice is a Microsoft open-source voice AI family with ASR and TTS tracks. The relevant transcription model is VibeVoice-ASR-7B, described by Microsoft as a unified speech-to-text model for 60-minute long-form audio in a single pass within a 64K-token context. Its output is structured around "who, when, what": speaker diarization, timestamps, and recognized content. It also supports customized hotwords for names, technical terms, and domain context. Source: microsoft/VibeVoice, 2026-07-01

The same repository also tracks long-form and realtime TTS: VibeVoice-TTS targets up to 90-minute conversational speech with up to four speakers, while VibeVoice-Realtime-0.5B targets streaming text input and roughly 300 ms first-audible latency. Keep those separate from ASR in notes and routing; the X bookmark was about transcription, but the source project is a broader voice model family. Source: microsoft/VibeVoice, 2026-07-01

Important caveat: Microsoft flags VibeVoice as research/development oriented and warns about synthetic speech risks. The repo notes that VibeVoice-TTS code was removed after out-of-scope use, while ASR remains the relevant local transcription candidate for Kevin's workflows. Source: microsoft/VibeVoice, 2026-07-01

Relevance to Kevin's Stack

Directly relevant to the voice-to-brain pipeline (recipes/voice-to-brain.md, scripts/voice-server.ts). Kevin's current voice integration uses OpenAI Realtime API for live calls - this Microsoft model could serve as a post-hoc transcription layer for recorded meetings, voice memos, or archived audio.

Also complements the meeting-sync pipeline (scripts/sync-meetings.ts) - Circleback handles transcription for meetings, but ad-hoc audio recordings (voice memos, recorded calls) currently have no local transcription path. VibeVoice-ASR is the candidate here, not VibeVoice-TTS. Any local adoption should run an eval on Kevin-style audio before trusting diarization or timestamps. Source: microsoft/VibeVoice, 2026-07-01

Local video review shows the referenced artifact as a project-page walkthrough: VibeVoice header, ASR/TTS badges, news about VibeVoice-ASR adoption, a VibeVoice-ASR-based input method, and code/demo frames. That matches the source repo better than the original generic "Microsoft transcription" page title. Source: local artifact review, 2026-07-01

Signal

  • ~6,988 combined engagement - strong developer interest in local long-form transcription

Timeline

  • 2026-04-08 | X bookmark amplifies VibeVoice as a Microsoft open-source voice AI that can transcribe 60 minutes of audio in one pass with speaker and timestamp structure. Source: X/@KanikaBK
  • 2026-05-17 | Added to wiki tool registry. Source: X bookmark absorption, 2026-05-17
  • 2026-07-01 | Source/artifact review corrected the page from a generic "Microsoft Voice Transcription" note into the VibeVoice family: ASR-7B for 60-minute speech recognition, TTS for long-form multi-speaker synthesis, and Realtime-0.5B for streaming TTS. Added local visual evidence and responsible-use caveat. Source: microsoft/VibeVoice, 2026-07-01

Related Pages

Voice-to-BrainMarkItDown - Universal File-to-Markdown ConverterSecurity and Review Skills