What problem does it solve? Transcribing long audio recordings with speaker labels and timestamps requires either expensive local GPU hardware or manual cloud setup. This Skill runs Microsoft's VibeVoice ASR models on Google Colab through the official Colab CLI, letting you choose between a low-cost CPU path and a high-accuracy GPU path without owning the hardware. ## Core Features & Use Cases - Dual model paths: Run the economical BitNet build (VibeASR.cpp, CPU-only) or the full 7B VibeVoice-ASR-HF model on A100/H100 GPUs, with an explicit, never-hidden fallback from full to BitNet. - Structured transcription output: Produces a plain transcript, a metadata JSON with model revision, hardware, duration, and parsed speaker/timestamp segments, plus a stderr log for auditing. - Hotwords and batch sessions: Pass domain vocabulary via --context (e.g., legal entity names) and reuse named Colab sessions to transcribe batches without recompiling or re-downloading models. - Use Case: A legal team needs to transcribe hours of Portuguese hearing recordings. Use the full model on an A100 with context hotwords like "IPERON, Sisprev, Procuradoria-Geral do Estado" to get accurate speaker-labeled transcripts with timestamps. ## Quick Start Use the vibevoice-asr skill to transcribe my recording.wav file on Colab with the full 7B model on an A100 GPU, including the hotwords IPERON and Sisprev.
