Offline Speech-to-Text Transcriber

The Core Problem: Balancing Security and Efficiency

Teams frequently need to convert recordings—whether from meetings, interviews, or online lectures—into text for summarization, archiving, or information retrieval. However, uploading files containing internal discussions, trade secrets, or unreleased data to third-party cloud services poses a significant data privacy risk. For enterprises, especially those in sectors with strict information security policies (like state-owned enterprises), this is a non-negotiable constraint. This skill provides a local-first, offline-capable transcription solution, ensuring raw audio data is processed exclusively on your local machine.

How It Works

The skill leverages local ASR (Automatic Speech Recognition) engines like whisperX. Its workflow involves these key steps:
1. Environment Check: It runs scripts/check_env.py to detect the readiness of core dependencies: ffmpeg, speech model libraries (e.g., whisperx), and optional speaker diarization models (e.g., pyannote). This check dictates the execution path.
2. Smart Branch Decision:
* Local First: If the local environment is ready, transcription is performed locally by default. This is the recommended path—the audio file never leaves your device.
* Cloud Fallback: Only when local engines are unavailable (e.g., missing dependencies that cannot be installed) and you explicitly consent to upload, will it call cloud APIs (like Alibaba Cloud's Paraformer or OpenAI Whisper). This is a degraded option; cloud providers typically do not support speaker separation.
3. Flexible Transcription: It supports common audio/video formats like m4a, mp3, wav, flac, and mp4. You can specify the model size (--model large-v3/medium/small) to trade off between accuracy and performance, and use the --hint-words flag to inject specific industry terms or names for better accuracy. For multi-speaker dialogues, the --diarize flag enables speaker diarization.
4. Standardized Output: By default, it generates timestamped .srt and .vtt subtitle files, along with a plain text .txt transcript. The output automatically includes a machine transcription disclaimer watermark, reminding downstream users to proofread the content.

Use Cases and Important Notes

• Not a Final Draft: ASR output is machine-generated and will contain errors, such as homophone substitutions or misrecognitions, especially with proper nouns, numbers, or accented speech. All transcribed text requires manual proofreading before use in formal minutes or publications.
• Privacy Safeguards: For sensitive scenarios in state-owned enterprises, the local-first mode is a hard requirement. Cloud functionality is an emergency fallback, and its activation requires explicit user confirmation.
• Integrated Workflows: The .txt and .srt outputs from this skill serve as standard inputs for downstream analysis skills (e.g., meeting minute generation, interview insight extraction). After transcription, you can be prompted to proceed to the next step in the workflow.