[Feature] Lightweight Model Option: Compact English/Light Model for Low-Resource Profiles

[SentimentalK]

1. Goal

Provide an optional compact, lightweight model (e.g. English-only or ultra-compact zipformer) to reduce RAM and disk footprint for memory-constrained environments, while keeping the high-quality bilingual X-ASR as the primary default.

2. Performance Profile Context

• Current Baseline (Quality / X-ASR):
  • RAM: ~700 MB (Model ~570 MB loaded into ONNX runtime)
  • CPU: ~1%–2% during active streaming
  • GPU: 0
  • Quality: Highly accurate bilingual Chinese + English.
• Target Baseline (Light Model Option):
  • Model file size: < 150 MB
  • Active RAM: <= 300 MB
  • CPU: < 1%
  • Use case: English-only dictation, low-memory laptops, background standby devices.

3. Architecture & Requirements

• Quality by Default: X-ASR remains the bundled, out-of-the-box default model.
• On-Demand Acquisition: Light model is acquired lazily when selected in settings.
• Registry Integration: Plugs into models/registry.json without altering the core pipeline.
• Benchmark Gate: Must pass accuracy and latency benchmarks before production release.

4. Acceptance Criteria

• User can select between "Quality (X-ASR)" and "Light (English)" in configuration/UI.
• Light model operates within <= 300MB RAM footprint.
• Switching models performs seamless transactional reload.
• Does not degrade out-of-the-box experience for default users.