zhipu-asr

What This Skill Does

The zhipu-asr skill is a powerful Automatic Speech Recognition (ASR) tool integrated into the OpenClaw agent ecosystem. Leveraging Zhipu AI's advanced GLM-ASR model, this skill allows users to convert spoken Chinese audio into highly accurate text. It is specifically designed to handle a variety of audio formats, automatically managing format conversion via ffmpeg. Whether you are dealing with short snippets or complex professional dialogues, this tool provides sophisticated features like context prompting and custom hotwords to ensure the transcript captures specialized terminology, names, and industry jargon with high precision.

Installation

To integrate this skill into your environment, use the OpenClaw package manager: clawhub install openclaw/skills/skills/franklu0819-lang/zhipu-asr

Ensure you have set your Zhipu AI API credentials in your shell environment: export ZHIPU_API_KEY="your-key-here"

Use Cases

• Meeting Transcription: Automatically transcribe recorded meeting audio, using previous segment summaries as context to maintain flow and terminology consistency.
• Technical Documentation: Transcribe voice notes taken during brainstorming sessions involving complex software architecture, using hotwords for specific library or project names.
• Medical or Professional Interviews: Utilize domain-specific hotwords to ensure that unique terminology, medical symptoms, or specific company jargon are recognized correctly by the AI model.
• Content Creation: Quickly convert spoken scripts, podcast segments, or interview recordings into editable text drafts for further content refinement.

Example Prompts

1. "Transcribe the file marketing_meeting.wav and use the hotwords 'Campaign, ROI, Quarterly' to ensure marketing terms are captured correctly."
2. "Process interview_part2.mp3. Use the following context: 'The previous discussion focused on the new Q3 product roadmap and the hiring strategy.'"
3. "Transcribe technical_demo.wav. Please pay special attention to the following technical terms: 'React, Kubernetes, Microservices, API.'"

Tips & Limitations

• Quality Control: While the skill handles various formats (WAV, MP3, OGG, M4A, AAC, FLAC, WMA), WAV at 16000Hz remains the optimal format for high-accuracy recognition.
• Constraints: Keep in mind that individual files must not exceed 25 MB or 30 seconds of duration. For longer files, split them into segments and utilize the context parameter to bridge the transcription continuity.
• Prompt Engineering: Use the 'hotwords' feature liberally when dealing with non-standard nouns, proper names, or proprietary industry acronyms to drastically improve accuracy on the first pass.