Speech-to-text service using GLM-ASR-Nano model with GPU management, supporting 17 languages and arbitrary-length audio with intelligent VAD segmentation
GLM-ASR has 5 tools with visible definitions in mcp_server.py. All tools have basic descriptions in Chinese and simple input schemas. However, critical quality issues prevent a higher score: (1) descriptions are minimal and lack context about when/why to use each tool; (2) parameter descriptions are present but sparse; (3) output schemas are completely undocumented, tools return dict/JSON but no schema is specified; (4) no error handling guidance, errors are returned but with no recovery hints for the LLM; (5) tool naming lacks consistency with verb_noun convention; (6) STDIO transport caps the entire server at 50 maximum, regardless of definition quality.
获取 GPU 和模型状态
加载模型到 GPU
重新加载模型
转录音频文件为文本
卸载模型,释放 GPU 显存
No output schemas documented for any tool. All tools return dicts but structure is inferred only from code, not specified in tool definitions. LLMs cannot plan downstream calls or extract fields with confidence.
Error handling is minimal. Tools catch exceptions and return {status: 'error', error: str(e)} but provide no recovery guidance. LLM receives raw exception messages without context on whether to retry, ask user, or escalate.
Descriptions are too brief (11-20 chars in English). Current descriptions fail to explain WHAT the tool does, WHEN to use it, or WHAT it returns.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 40 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 41 | - | v1 |
Parameter constraints and formats are missing. 'audio_path' has no validation rules (supported formats, max size), 'checkpoint' has no enum of valid model IDs or pattern. LLMs cannot self-correct invalid inputs.
STDIO transport only. Server is not remotely accessible and cannot be used by hosted MCP clients. Hard cap of 50 on protocol readiness score.