feat: speech-to-text NVIDIA/non-NVIDIA split — CPU-capable alternative for laptops

[LTSCommerce]

Problem

play-speech-to-text.yml currently installs faster-whisper with CUDA support, which is only useful on machines with a discrete NVIDIA GPU. On laptops without NVIDIA (integrated Intel/AMD graphics only), CUDA is unavailable and the CUDA backend is pointless — wastes install time and may produce confusing errors or silent fallback to CPU anyway.

Current Behaviour

Single installation path: faster-whisper + CUDA regardless of hardware.

Expected Behaviour

Detect whether NVIDIA GPU is present and choose the appropriate backend:

• NVIDIA GPU detected → faster-whisper with CUDA (current behaviour, keep as-is)
• No NVIDIA GPU → CPU/ROCm-compatible alternative

Investigation Needed

Research and evaluate speech-to-text systems that work well on CPU or integrated graphics:

• faster-whisper with CPU backend — same tool, just skip CUDA deps; check if performance is acceptable on modern laptop CPUs
• whisper.cpp — pure C++ implementation, no Python deps, runs well on CPU, supports Metal/OpenCL
• vosk — lightweight offline STT, very low resource usage, runs on CPU
• RealtimeSTT with CPU — check if the current wsi-stream wrapper can work without CUDA
• sherpa-onnx — ONNX-based, good CPU performance, supports Whisper models

Acceptance Criteria

• Playbook detects NVIDIA GPU (reuse logic from check_hardware / lspci | grep -i nvidia)
• NVIDIA path: faster-whisper + CUDA (current)
• Non-NVIDIA path: working alternative with acceptable latency on laptop CPU
• Both paths use the same wsi-stream interface (or document differences)
• play-speech-to-text.yml removed from auto_run_common in run.bash until this is resolved (currently auto-runs on all hardware)

Related

• playbooks/imports/optional/common/play-speech-to-text.yml
• files/home/.local/bin/wsi-stream
• run.bash auto_run_common array