Static Recorder Pro 发布：浏览器内录音、Whisper 转写与 AI 纪要

Record calls, get accurate transcripts, and draft AI notes — in your browser, on your machine, with your choice of AI.

Static Recorder Pro is a fully self-contained, single-file web app. It records call audio (microphone + system audio simultaneously), transcribes it locally with Whisper, and drafts notes, replies, and follow-ups — either with a local in-browser LLM or (optionally) a third-party cloud LLM API of your choice. No backend is required; everything runs from a static page.

What it does

Local-First, With an Optional Cloud Mode

The AI Co-Pilot has two modes, switched with a tab at the top of the AI panel:

Local (In-Browser) — default

• Entirely offline after the one-time model download; no API keys, no accounts, no data leaves your device.
• Runs Whisper + a small LLM through Transformers.js on WebGPU (falling back to WASM/CPU).
• Private and free, but constrained by browser memory and the size of small local models (see Known limitations).

Cloud API — optional

• Bring your own key from OpenAI, Anthropic (Claude), Google Gemini, Groq, OpenRouter, or a Custom OpenAI-compatible endpoint (LM Studio, Ollama, enterprise gateways, etc.).
• Per-provider model dropdowns (e.g. gemini-3.8-flash, claude-sonnet-4-5, gpt-4o-mini), or type any model in Custom.
• Responses stream token-by-token; the same prompt instruction, format reference, compaction, and presets apply.
• No browser memory limits — full 128k+ context handling of long calls.

Note: the API key is stored only in the page's localStorage and is sent directly from your browser to the provider. Use a restricted/limited key — anyone with access to the device can read it.

Providers that are CORS-enabled (work directly from a static page): OpenAI, Anthropic (via its browser-access header), Google Gemini, Groq, OpenRouter, and Custom endpoints that send CORS headers. OpenCode Go is notably excluded — its API sends no CORS headers, so it cannot be called from a static page; it must be used from the OpenCode app itself.

Models

Local transcription (Whisper)

Local AI notes (in-browser LLM)

Local models are downloaded from Hugging Face once and cached by the browser; every run afterward works fully offline.

Cloud AI notes

Model lists are populated per provider (e.g. gemini-3.8-flash for Gemini, gpt-4o-mini for OpenAI, claude-sonnet-4-5 for Anthropic). If your saved model isn't in the list, it appears as an extra (custom) option so nothing is lost.

Tech stack

How to use it

1. Open the app

Open index.html in any modern browser (Chrome, Edge, Firefox — Chrome/Edge recommended for WebGPU). It works from file://, GitHub Pages, or any static host.

2. Record a call

1. Press Start Recording.
2. Choose your microphone, then share the call tab/window and include system audio (so the Caller's voice is captured).
3. Watch the live waveform meters, then press Stop — the recording saves locally instantly.

3. Transcribe

1. Click Transcribe Audio.
2. Whisper processes the audio in chunks with live, streaming output.
3. You get a timestamped, speaker-labeled transcript ([Agent] / [Caller]).

4. Generate AI notes (Local mode)

1. The transcript appears in the AI Assistant box.
2. Choose Local (In-Browser) (default), pick a model + output style.
3. (Optional) type a prompt instruction, e.g. "Summarize this call in 5 bullets."
4. Press Generate with AI — responses stream in token-by-token.

5. Generate AI notes (Cloud mode)

1. Switch to the Cloud API tab.
2. Pick a provider (e.g. OpenAI), choose its model, and paste your API key (saved in this browser only).
3. For Custom, enter your base URL (e.g. https://my-endpoint.example/v1) and model name.
4. Press Generate with AI — the request goes straight from your browser to the provider and streams back.

6. Note templates & formatting

• Use the Format Reference / Template box to show the AI the layout you want (ConnectWise notes, emails, etc.).
• The small toolbar (B / I / U / H / • / 1.) inserts formatting into the template: **bold**, *italic*, ## headers, and lists.
• The AI output renders that formatting live, and Copy AI Text (Rich) pastes it into ConnectWise/Word/email with the formatting intact.

7. Custom prompt & reference presets

• Click the + button next to the AI Prompt Instruction chips to save your own prompt as a reusable chip.
• Do the same for Format References — both persist across visits.
• Right-click a custom chip to delete it.

8. Manage history

The Call History tab lets you: listen to tracks, Save Agent/Caller (.wav), Save Transcript (.txt), re-transcribe, Use Transcript (load into the AI), AI Draft (load + generate immediately), refresh, or Clear All — all stored locally in IndexedDB.

Fix slow local AI generation

Local AI generation is dramatically faster on the GPU. A badge in the top-right of the screen shows the mode:

• WebGPU Active — GPU accelerated (fast)
• WASM CPU Mode — CPU (works, slower)
• Cloud API Mode — using your configured cloud provider

If you see WASM CPU Mode and want WebGPU:

1. Open chrome://flags/#enable-vulkan and set Vulkan to Enabled.
2. (Optional, to be safe) enable:
   • chrome://flags/#enable-unsafe-webgpu
   • chrome://flags/#ignore-gpu-blocklist
3. Restart Chrome.
4. Reload — the badge should read WebGPU Active.

Notes:

• WebGPU requires Chrome or Edge. Firefox/Safari always fall back to CPU (everything still works, just slower).
• Keep hardware acceleration on: Settings → System → "Use graphics acceleration when available."
• On Linux, the Vulkan flag is the usual missing piece — Chrome ships it disabled by default.
• If a local model runs out of memory, the app automatically retries with the next smaller model, and switches to CPU if WebGPU execution fails.

Smart transcript compaction

When you press Generate with AI, the transcript is compacted in the background (the visible box stays unchanged):

[00:00:01.290 - 00:00:07.930] [Caller] "Well the issue with our system"
[00:00:07.930 - 00:00:11.210] [Caller] "we think is related to memory."

becomes:

[Caller] "Well the issue with our system we think is related to memory"

Timestamps are removed and consecutive same-speaker lines are merged — typically ~45% fewer characters, so the AI reads more of the call within its context budget.

Tips & notes

• First local run downloads models. Allow time; they're cached afterwards.
• Long calls + local mode: browser memory limits the local path. Keep transcripts compact (the app does this automatically) or use Cloud mode, which handles full-length calls easily.
• Cloud keys are stored in the browser — use a restricted key with a spend cap.
• OpenCode Go isn't supported from a static page (no CORS); use it with the official OpenCode app.
• Choose the right local model for your machine. SmolLM2 360M for CPU-only machines; Qwen2.5 on WebGPU for best quality.

Privacy & security

• Local mode: zero network transmission of your audio or text (except one-time model downloads from Hugging Face).
• Cloud mode: your transcript, prompt, reference, and key are sent directly from your browser to the provider you configure — with no intermediaries in between, but the provider's terms apply.
• No telemetry, no analytics, no cookies.
• Recordings, transcripts, and settings live in your browser's local storage. Clearing site data erases them permanently.

Requirements

Known limitations

• Small local models are private but less capable than cloud LLMs — they can simplify nuance or follow formatting loosely.
• Local mode memory ceiling: WebAssembly is 32-bit (~4 GB heap cap), so long conversations in local mode can hit ONNX "tensor too large" / out-of-memory errors. The app auto-falls back to CPU and smaller models, but Cloud mode is the reliable route for long calls.
• Cloud mode requires CORS-enabled providers — OpenAI, Anthropic, Gemini, Groq, OpenRouter, and custom CORS-enabled endpoints work; others (e.g. OpenCode Go) don't.
• Trimming/compaction helps the AI, but the visible transcript is never modified without your say-so.

License

Provided for personal/learning use. Local models are covered by their respective licenses (Whisper + ONNX repos — see Hugging Face). Cloud usage is subject to your provider's terms.