No darwin/arm64 images for sherpa-onnx and speaker-recognition backends — diarization and /v1/voice/embed unavailable on macOS

[enneo-service]

LocalAI version: 4.4.1 (Homebrew bottle)
Environment: macOS (Apple Silicon, M1 Max), darwin/arm64

What happened

Neither the sherpa-onnx nor the speaker-recognition backend has a darwin/arm64 image, which leaves the new audio-identity endpoints without any working backend on macOS:

$ local-ai backends install sherpa-onnx
failed to get image "quay.io/go-skynet/local-ai-backends:latest-cpu-sherpa-onnx":
no child with platform darwin/arm64 in index quay.io/go-skynet/local-ai-backends:latest-cpu-sherpa-onnx

$ local-ai backends install speaker-recognition
failed to get image "quay.io/go-skynet/local-ai-backends:latest-cpu-speaker-recognition":
no child with platform darwin/arm64 in index quay.io/go-skynet/local-ai-backends:latest-cpu-speaker-recognition

Consequence

• /v1/audio/diarization has no usable backend on macOS: the sherpa-onnx path is unavailable (this issue), and the vibevoice-cpp darwin image is broken separately (darwin/arm64 vibevoice-cpp backend image ships sources instead of build artifacts — backend cannot start #10267).
• /v1/voice/embed is unavailable on macOS entirely.

Request

Publish darwin/arm64 children for the sherpa-onnx and speaker-recognition backend images. sherpa-onnx is ONNX/CPU-based, so a darwin build should be reasonably portable — and it would make the diarization stack usable for the (large) population of local-AI users on Apple Silicon.

For context: metal-llama-cpp and metal-whisper install and run great on the same machine — LLM + ASR on macOS are already in excellent shape.

[localai-bot]

Fixed by #10275 (merged).

Both backends were simply absent from the darwin/arm64 build matrix and the gallery had no metal: capability for them. What changed:

• Added sherpa-onnx (Go, metal) and speaker-recognition (Python, mps) to the darwin build matrix.
• Wired the metal: capability + metal-* image aliases into backend/index.yaml for both.
• Added a requirements-mps.txt for speaker-recognition so its mac build installs plain onnxruntime (which has a macOS arm64 wheel) instead of onnxruntime-gpu (which doesn't).

Verification:

• speaker-recognition: darwin/arm64 build verified green on the macOS CI runner.
• sherpa-onnx: its Makefile already fetches the onnxruntime-osx-arm64 runtime, so the build is expected to be portable; it'll be exercised on the master backend-image build.

The darwin/arm64 child images become available once the backend images are published from master. After that, local-ai backends install sherpa-onnx and local-ai backends install speaker-recognition should resolve the darwin/arm64 platform, and /v1/audio/diarization (sherpa-onnx path) + /v1/voice/embed will have a working backend on Apple Silicon. Please re-test once the refreshed images land — and shout if sherpa-onnx misbehaves on macOS.