Problem
Kastword offers speech-model downloads ranging from roughly 78 MB to 3.1 GB, but the model manager describes them only with qualitative speed and accuracy labels. It checks available disk space before downloading, yet users receive no memory, expected-performance, or hardware-suitability guidance before selecting a potentially expensive or unusable model.
Expected outcome
Users can choose a model likely to work acceptably on their machine and understand the speed, accuracy, storage, and memory trade-offs before downloading it.
Scope
- Add conservative, version-controlled resource metadata for every model in the supported catalog.
- Show approximate memory and storage requirements alongside the existing speed and accuracy characteristics.
- Explain the trade-offs behind recommended models.
- Where reliable platform information is available, warn before downloading or activating a model that appears unsuitable for the available resources.
- Clearly present estimates as approximate because runtime behavior varies by backend, recording, and hardware.
Non-goals
- Benchmarking every machine.
- GPU acceleration.
- Quantized, distilled, or community models.
- Promising exact transcription latency.
- Blocking a user choice solely because performance may be slow.
Acceptance criteria
- Every catalog entry displays approximate memory and storage requirements.
- Recommendations explain the speed, accuracy, and resource trade-off.
- Clearly unsuitable selections receive a warning before download or activation.
- Warnings remain advisory unless safe execution cannot be guaranteed.
- Missing or unreliable system resource information is handled without presenting false certainty.
- Resource metadata is version-controlled and validated with the model catalog.
- Deterministic tests cover metadata validation, display, recommendation boundaries, warning behavior, and unavailable resource information.
- User-facing documentation explains that actual resource use and transcription speed vary.
Dependencies and risks
Memory estimates vary by model, backend, architecture, and workload. Guidance must therefore be conservative, testable, and explicitly approximate.
Related work
Issue #15 added the model manager, download sizes, and qualitative model characteristics but did not address runtime suitability.
Problem
Kastword offers speech-model downloads ranging from roughly 78 MB to 3.1 GB, but the model manager describes them only with qualitative speed and accuracy labels. It checks available disk space before downloading, yet users receive no memory, expected-performance, or hardware-suitability guidance before selecting a potentially expensive or unusable model.
Expected outcome
Users can choose a model likely to work acceptably on their machine and understand the speed, accuracy, storage, and memory trade-offs before downloading it.
Scope
Non-goals
Acceptance criteria
Dependencies and risks
Memory estimates vary by model, backend, architecture, and workload. Guidance must therefore be conservative, testable, and explicitly approximate.
Related work
Issue #15 added the model manager, download sizes, and qualitative model characteristics but did not address runtime suitability.