docs(frontend): document local FunASR transcription

autogpt_platform/frontend/.env.default

@@ -82,6 +82,11 @@ NEXT_PUBLIC_VAPID_PUBLIC_KEY=BBg49iVTWthVbRYphwmZNvZyiSJDqtSO4nmLxDzLKe3Oo9jbtu0
 # OpenAI (for voice transcription)
 OPENAI_API_KEY=
 
+# Optional OpenAI-compatible voice transcription endpoint
+#TRANSCRIPTION_API_BASE_URL=
+#TRANSCRIPTION_API_KEY=
+#TRANSCRIPTION_MODEL=
+
 # Voice brain-dump onboarding step (`onboarding-brain-dump` LaunchDarkly
 # flag). Commented deliberately: an active value here would ship the feature
 # on to every install. Uncomment in your own .env, pair it with

autogpt_platform/frontend/src/app/api/transcribe/__tests__/route.test.ts

@@ -43,7 +43,7 @@ describe("transcribe route", () => {
 
   it("sends audio to a configured OpenAI-compatible transcription endpoint", async () => {
     process.env.TRANSCRIPTION_API_BASE_URL = "http://funasr.local:8000/v1/";
-    process.env.TRANSCRIPTION_MODEL = "iic/SenseVoiceSmall";
+    process.env.TRANSCRIPTION_MODEL = "sensevoice";
     const fetchMock = vi.fn().mockResolvedValue(
       new Response(JSON.stringify({ text: "hello from funasr" }), {
         status: 200,
@@ -68,7 +68,7 @@ describe("transcribe route", () => {
     expect(fetchInit.headers).toBeInstanceOf(Headers);
     expect((fetchInit.headers as Headers).has("Authorization")).toBe(false);
     const upstreamBody = fetchInit.body as FormData;
-    expect(upstreamBody.get("model")).toBe("iic/SenseVoiceSmall");
+    expect(upstreamBody.get("model")).toBe("sensevoice");
     const file = upstreamBody.get("file") as File;
     expect(file.name).toBe("recording.webm");
   });

docs/platform/copilot-local-llm.md

@@ -5,6 +5,8 @@
 > Generator block (used inside agent graphs you build yourself), see
 > [Running Ollama with AutoGPT](ollama.md). The two paths read different
 > env vars, so configuring one does not configure the other.
+> The optional voice transcription section below is also separate and
+> reads server-side variables from the frontend environment.
 >
 > Self-hosting only — the cloud `agpt.co` deployment routes AutoPilot
 > through Anthropic / OpenRouter and ignores the variables below.
@@ -96,6 +98,78 @@ CHAT_FAST_STANDARD_MODEL=hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
 CHAT_FAST_ADVANCED_MODEL=qwen3:14b-q4_K_M
 ```
 
+## Optional: local OpenAI-compatible voice transcription
+
+AutoPilot's microphone input can use any service that exposes the
+OpenAI-compatible `POST /v1/audio/transcriptions` route. Configure the
+base URL and model name for your service. The example below uses
+[FunASR](https://github.com/modelscope/FunASR), a self-hosted server that
+uses SenseVoice by default.
+
+Install the example server and start it on the Docker host:
+
+```bash
+python -m pip install -U funasr fastapi uvicorn python-multipart
+
+# CPU
+funasr-server --model sensevoice --device cpu
+
+# Or, on a CUDA host
+funasr-server --model sensevoice --device cuda
+```
+
+The first start downloads the configured model. The server listens on
+`0.0.0.0:8000` and does not require an API key by default. Before
+connecting AutoGPT, verify it with a local audio file:
+
+```bash
+curl http://localhost:8000/v1/audio/transcriptions \
+  -F file=@sample.wav \
+  -F model=sensevoice
+```
+
+Keep port 8000 on a trusted host or private network. If the endpoint is
+reachable from an untrusted network, put it behind an authenticated
+HTTPS reverse proxy. For another compatible service, substitute its
+startup command, port, and model name; the AutoGPT configuration below
+is otherwise the same.
+
+Then add the transcription settings to
+`autogpt_platform/frontend/.env`:
+
+```bash
+# Docker Desktop on macOS or Windows
+TRANSCRIPTION_API_BASE_URL=http://host.docker.internal:8000/v1
+TRANSCRIPTION_MODEL=sensevoice
+TRANSCRIPTION_API_KEY=
+```
+
+On native Linux Docker, replace `host.docker.internal` with the host's
+LAN or bridge-reachable IP. Alternatively, add
+`host.docker.internal:host-gateway` to the frontend service's
+`extra_hosts`. For frontend development outside Docker, use
+`http://localhost:8000/v1`.
+
+Recreate the frontend container so the server-side transcription route
+receives the new environment:
+
+```bash
+cd autogpt_platform
+docker compose up -d --force-recreate frontend
+```
+
+Leave `TRANSCRIPTION_API_KEY` empty when the transcription service does
+not require authentication. Set it when the endpoint or a reverse proxy
+requires bearer authentication. If no `TRANSCRIPTION_*` overrides are
+set, AutoGPT uses `whisper-1` and sends requests to
+`OPENAI_API_BASE_URL` when configured, or to the default OpenAI endpoint
+otherwise. `OPENAI_API_KEY` is attached only when the resulting base URL
+is the default OpenAI URL. Set `TRANSCRIPTION_API_KEY` explicitly when a
+custom `OPENAI_API_BASE_URL` requires bearer authentication.
+
+For the example server's model choices and deployment options, see the
+[model selection guide](https://github.com/modelscope/FunASR/blob/main/docs/model_selection.md).
+
 ## Picking a model
 
 The platform's chat loop calls **OpenAI-style tool-calling** on every