菜单
Voice notes are transcribed to text before the agent runs (local Python, OpenAI-compatible cloud, Gemini, or Hugging Face Inference API).
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Avishai-Tsabari/mercury --skill voice-transcribe命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Use this skill whenever the user wants to do anything with PDF files -- reading, extracting text/tables, merging, splitting, rotating, watermarking, creating, filling forms, encrypting/decrypting, extracting images, OCR, or converting to images. Activate when a .pdf file is mentioned or PDF output is requested.
Call TradeStation REST API v3 from the agent (accounts, balances, positions, bars) and place orders via mrctl two-step confirmation. Admin-only; requires Mercury host OAuth setup and token refresh job.
Access Gmail, Google Drive, Google Calendar, Docs, and Sheets via the gws CLI. Use for email (read, send, triage inbox, reply, forward), calendar (today's agenda, add or reschedule events, meeting prep), Drive (list, find, organize files), and Docs/Sheets (read, summarize, append). Triggers on any Google Workspace or Gmail request.
基于 SOC 职业分类
正在显示 SKILL.md
| name | voice-transcribe |
| description | Voice notes are transcribed to text before the agent runs (local Python, OpenAI-compatible cloud, Gemini, or Hugging Face Inference API). |
When a user sends a voice note or audio attachment, Mercury runs the voice-transcribe extension before the container starts. The transcript is appended to the user message as a [Voice transcript] block. You receive normal text — you do not need to read or play audio files for intent.
Keys resolve per space, then fall back to deployment-wide defaults (@global scope set from the dashboard Features page, or the extensions: section of mercury.yaml), then to the extension defaults.
| Key | Purpose |
|---|---|
voice-transcribe.provider | local (default), openai, gemini, or api |
voice-transcribe.model | Model id for the chosen provider (see below) |
voice-transcribe.local_engine | Local only: transformers (default) or faster_whisper (CTranslate2 Hub repos) |
voice-transcribe.base_url | openai only: API root (default https://api.openai.com/v1; Groq: https://api.groq.com/openai/v1) |
voice-transcribe.language | Cloud providers: ISO-639-1 hint (e.g. he). Improves accuracy on short voice notes |
local)Runs scripts/transcribe.py on the Mercury host.
Transformers (default) — voice-transcribe.local_engine=transformers uses Transformers pipeline("automatic-speech-recognition", model=...). Use standard PyTorch Whisper / compatible Hub ids (e.g. mike249/whisper-tiny-he-2).
Faster-Whisper — set voice-transcribe.local_engine=faster_whisper and faster-whisper loads CTranslate2 checkpoints from the Hub. Set voice-transcribe.model to a CT2 repo (not plain PyTorch Whisper weights). Examples for Hebrew: ivrit-ai/faster-whisper-v2-d4, ivrit-ai/whisper-large-v3-turbo-ct2.
Install on the same machine that runs Mercury:
pip install -r /path/to/.mercury/extensions/voice-transcribe/requirements.txt
Or: pip install "transformers>=4.40.0" "torch>=2.0.0" "faster-whisper>=1.0.0" "imageio-ffmpeg>=0.4.9"
Telegram voice is usually .ogg (Opus). Both backends shell out to an executable named ffmpeg; transcribe.py uses imageio-ffmpeg’s binary and exposes it as ffmpeg / ffmpeg.exe on PATH (hardlink or copy under %TEMP%\\mercury-voice-ffmpeg on Windows). Installing system ffmpeg on PATH also works (e.g. winget install Gyan.FFmpeg).
Optional host .env:
MERCURY_VOICE_PYTHON — Python executable (default: python on Windows, python3 elsewhere)MERCURY_VOICE_TRANSCRIBE_TIMEOUT_MS — subprocess timeout in ms (default 300000)MERCURY_VOICE_ASR_DEVICE — cpu, cuda, or auto. Used by both local engines. On Windows the default is CPU (CUDA often looks available but fails at inference). Set cuda if you have a working GPU stack.MERCURY_VOICE_FW_COMPUTE_TYPE — Faster-Whisper only: e.g. int8, float16, float32. If unset: int8 on CUDA, float32 on CPU.MERCURY_VOICE_LANGUAGE — Faster-Whisper only: ISO code passed to transcribe(..., language=...) (e.g. he). If unset, language is auto-detected.Hub messages like Xet Storage… hf_xet are warnings only, not the cause of failures.
First run downloads the model into the Hugging Face cache (size depends on the model). Each voice note spawns a new Python process, so the first transcription after Mercury starts can be slow while the model loads (no in-process reuse yet).
openai)POSTs audio to {base_url}/audio/transcriptions (multipart). Requires MERCURY_STT_API_KEY on the host (the key never enters agent containers). Works with:
base_url) — models gpt-4o-mini-transcribe (default), gpt-4o-transcribe, whisper-1voice-transcribe.base_url=https://api.groq.com/openai/v1, model whisper-large-v3No Python, ffmpeg, or GPU on the host — ideal for small VPS deployments. Set voice-transcribe.language (e.g. he) for better accuracy on short notes.
gemini)Sends audio inline to the Gemini API (generateContent) with a transcription instruction. Requires MERCURY_STT_GEMINI_API_KEY on the host (a plain API key — no GCP project setup; deliberately separate from the model-chain MERCURY_GEMINI_API_KEY, which must keep flowing into containers). Default model gemini-2.5-flash. Also host-only, no local dependencies.
api)POSTs audio to https://api-inference.huggingface.co/models/<model>. Requires MERCURY_HF_TOKEN on the host. Pick a model that has an Inference Provider on its Hub page (e.g. openai/whisper-large-v3). Hebrew-tuned models such as mike249/whisper-tiny-he-2 often have no hosted provider — use local for those. CT2 / Faster-Whisper Hub repos are for the local engine only, not this API.
Only callers with the voice-transcribe permission (default: admin + member) get transcription. Others keep the raw message without an appended transcript.