docs: add FunASR SenseVoice STT setup

[LauraGPT]

Summary

• Document FunASR/SenseVoice through LibreChat's existing OpenAI-compatible STT integration, including service setup, endpoint checks, separate Python environment, matching Torch/torchaudio builds, and network-access precautions.
• Synchronize with current docs main, including the Next.js 16.2.12 and pnpm 9.15.9 updates.
• Fix the self-hosted example for the current speech SSRF contract: add an exact speech.stt.allowedAddresses entry for the configured private endpoint. Without it, a private/loopback speech destination is blocked.
• Explain host-process, Docker-host and Compose deployment choices by pairing the URL hostname with its exact host-and-port exemption. Keep only the required entry; do not disable the guard or exempt an entire network. Link to the upstream speech SSRF documentation.
• Keep deployment values readable on mobile by stating the shared URL scheme, port and path once. The full YAML example and service commands remain available unchanged except for the required exemption.

Fresh Validation (2026-09-29)

Signed and DCO-compliant head: 74c442b6946ec76741cb7973130a2c99671ff335.
Merged actual main at c7ab51f1e92bbf4eaf4e4d7635c208cc589593b3; the previous PR head was 34 commits behind it. Freshness is checked against branches/main, not the PR API's base snapshot. Effective PR scope remains only content/docs/configuration/stt_tts.mdx.

Validation uses Node 24.20.0, repository-pinned pnpm 9.15.9 and the frozen 871-package dependency graph:

• pnpm install --frozen-lockfile, native pnpm lint, pnpm typecheck and target-document Prettier check passed.
• pnpm test: 34 test files, 404 tests passed on the final candidate.
• Separate structured MDX/YAML controls: 5 passed. Before the exemption fix, the exact-endpoint and deployment-guidance checks failed while three preservation checks passed. All shell blocks, other YAML examples and the existing SenseVoice endpoint/model settings remain unchanged from the merged baseline.
• CIRCLE_NODE_TOTAL=3 pnpm build: 2,911 static pages, native postbuild and sitemap generation completed on the final candidate. Next.js used two workers; no project worker configuration was changed. Final build: 260.94 seconds. Its Fumadocs dynamic-import cache-dependency warnings do not represent a failed build.
• Production route /en/docs/configuration/stt_tts: HTTP 200 at 1440x1000 and 390x844 with Playwright 1.58.2 and its matching Chromium 145.0.7632.6. Checks covered the configuration, all three deployment address pairs, unclipped inline deployment values, no page-level horizontal overflow or page script errors, and the linked SSRF section returning HTTP 200 with the expected anchor. Final desktop/mobile screenshots were visually inspected. External browser requests were blocked and non-essential analytics declined in the temporary test contexts.
• PR-relative git diff --check passed. The native pre-commit hook ran Prettier and ESLint on all 90 matching staged code files inherited from the merge; it left the tested tree unchanged. No hook was bypassed.
• Independent read-only review of the final document found no introduced P1/P2. Final document SHA-256: 4ecad23e09884990a35af140c544e9510b5686e1fa296e46d8ff61f9e32ab6eb.

Broad checks used an isolated export of Git-tracked sources, excluding unrelated pre-existing untracked task files. The committed document matches the built and reviewed document byte-for-byte. The initial interactive dependency command exited without installing; it was not counted as success and was followed by a completed noninteractive frozen installation and normal native hook setup.

This is documentation/site validation, not a new FunASR clean installation, model inference, real DNS/Docker connectivity, authentication acceptance, or end-to-end LibreChat recording test. The speech SSRF contract was checked against current upstream documentation and test source, not claimed as an executed LibreChat backend suite. No speech model, GPU workload or paid API was run, and no deployment approval was granted.

Remote readback verified the signature, both parents, tree and document bytes. A fresh comparison against actual branches/main reported 3 ahead / 0 behind.

Hosted Checks

On this exact head, Vercel reports failure: Authorization required to deploy. Authorization is required from the LibreChat Vercel team. GitHub CI 36506893696 and Next.js Bundle Analysis 36506893846 are completed / action_required, with no check runs. These are approval holds, not running or passing hosted checks. Local production validation does not grant or replace approval; no deployment approval or bypass was attempted.

Historical Model Smoke Test

The earlier head 4e726cbf645c329e74cbbbe1af14a4fa63c31adc was tested with FunASR 1.3.27 on CPU: /health returned HTTP 200, and a LibreChat-shaped multipart request (file, model=sensevoice, language=zh) returned HTTP 200 with a Chinese transcript in 0.574s.

That real-model test was not rerun for this documentation-only follow-up. The fresh website tests do not establish a new clean FunASR installation, model inference result, or end-to-end LibreChat recording session.

[vercel]

@LauraGPT is attempting to deploy a commit to the LibreChat's projects Team on Vercel.

A member of the Team first needs to authorize it.

[LauraGPT]

@berry-13 This FunASR/SenseVoice STT docs integration is now rebased onto current main at signed head 4e726cbf. Exact-head validation passes lint, 184 tests, typecheck, Prettier, the full 2,877-page production build, the rendered docs route, and a real FunASR 1.3.27 CPU transcription. The sole red status is Vercel team authorization, not a build failure. Could you review it when convenient?