Skip to main content
SlasshyWispr analyzes your hardware to recommend optimal local models for speech-to-text and AI inference.

​
Hardware Advisor

SlasshyWispr includes an intelligent hardware advisor that analyzes your system and provides model recommendations.

​
What Gets Analyzed

The hardware advisor collects: 
interface LocalSttHardwareAdviceResponse {
  cpuName: string;                    // CPU model name
  logicalCores: number;               // Number of logical CPU cores
  totalRamGb: number;                 // Total system RAM in GB
  nvidiaGpuDetected: boolean;         // NVIDIA GPU presence
  gpuName: string;                    // GPU model name
  gpuVramGb: number;                  // GPU VRAM in GB
  performanceTier: string;            // Overall system tier
  slasshySuggestionModel: string;     // Top recommended model
  suggestedModels: string[];          // All recommended models
  cautionModels: string[];            // Models that may struggle
  selectedModelWarning: string;       // Warning for current selection
  details: string;                    // Additional information
}
The hardware advisor runs automatically when you first access local model settings. Results are cached to avoid repeated system scans.

​
CPU Requirements

​
Minimum Requirements

  • Processor: Modern x86_64 or ARM64 CPU
  • Cores: 2 logical cores minimum
  • Instruction Sets: AVX support recommended for optimal performance

Entry Level

4-6 coresIntel Core i3/i5, AMD Ryzen 3/5, Apple M1Suitable for: Moonshine, SenseVoice, Whisper Small

Mid Range

6-8 coresIntel Core i5/i7, AMD Ryzen 5/7, Apple M2/M3Suitable for: Parakeet, Whisper Medium, small Ollama models

High End

8+ coresIntel Core i7/i9, AMD Ryzen 7/9, Apple M2/M3 Pro/MaxSuitable for: All models, large Ollama models

​
CPU Performance Impact

  • STT Transcription: CPU-intensive for model inference
  • Ollama Inference: Benefits greatly from more cores
  • Concurrent Operations: Multiple cores enable simultaneous STT + AI
Apple Silicon Macs (M1/M2/M3) provide excellent performance due to unified memory architecture and Neural Engine acceleration.

​
RAM Requirements

​
Memory Guidelines

RAM requirements depend on model sizes:
Total RAMSTT ModelsOllama Models
4 GBMoonshine onlyNot recommended
8 GBMoonshine, SenseVoice, Whisper Small1B-3B models
16 GBAll small/medium models, Parakeet7B-8B models
32 GBAll models including Whisper Turbo13B-14B models
64+ GBAll models with headroom30B+ models

​
RAM Usage Patterns

1

Base Application

SlasshyWispr uses ~200-400 MB for the core application
2

STT Model Loading

Models consume RAM approximately equal to their file size:
  • Moonshine: ~60 MB
  • SenseVoice: ~160 MB
  • Parakeet/Whisper Small/Medium: ~500 MB
  • Whisper Large/Turbo: 1.1-1.6 GB
3

Ollama Model Loading

LLM models consume significant RAM:
  • 1B-3B models: 2-4 GB
  • 7B-8B models: 6-10 GB
  • 13B-14B models: 12-18 GB
  • 30B+ models: 25+ GB
4

Operating System Reserve

Keep 2-4 GB free for OS and other applications
Insufficient RAM will cause severe performance degradation through swap/paging. Always choose models that fit comfortably within your available RAM.

​
GPU Support

​
NVIDIA GPU Detection

SlasshyWispr automatically detects NVIDIA GPUs and reports:
  • GPU model name
  • VRAM capacity
  • CUDA availability

​
GPU Acceleration Benefits

STT Acceleration

NVIDIA GPUs can accelerate ONNX model inferenceSpeed improvement: 2-5x faster transcription

Ollama Acceleration

Ollama automatically uses GPU for LLM inferenceSpeed improvement: 5-20x faster responses

​
Supported GPUs

NVIDIA (CUDA):
  • GTX 1000 series and newer
  • RTX 2000, 3000, 4000 series
  • Professional GPUs (Quadro, Tesla, A-series)
  • Requires: CUDA 11.x or newer
Apple (Metal):
  • M1, M2, M3 series (integrated GPU)
  • Metal acceleration automatic on macOS
AMD (ROCm):
  • Limited support via Ollama on Linux
  • Not supported for ONNX Runtime in SlasshyWispr

​
VRAM Requirements

VRAMSTT ModelsOllama Models
2-4 GBSmall models1B-3B models
6-8 GBMedium models7B-8B models
12 GBLarge models13B-14B models
16+ GBAll models30B+ models
If VRAM is insufficient, models will fall back to CPU/RAM. This is slower but still functional.

​
Performance Tiers

Based on your hardware analysis, SlasshyWispr assigns a performance tier:

​
Tier 1: Entry Level

Hardware:
  • 4-8 GB RAM
  • 2-4 CPU cores
  • No dedicated GPU
Recommended Models:
  • STT: Moonshine Base, SenseVoice
  • AI: Ollama not recommended, use online mode
Use Case: Basic dictation with cloud AI

​
Tier 2: Standard

Hardware:
  • 8-16 GB RAM
  • 4-6 CPU cores
  • Optional: Entry-level GPU (2-4GB VRAM)
Recommended Models:
  • STT: Parakeet v3, Whisper Small/Medium
  • AI: Ollama 1B-3B models (llama3.2:1b, mistral:3b)
Use Case: Balanced offline performance

​
Tier 3: Performance

Hardware:
  • 16-32 GB RAM
  • 6-8 CPU cores
  • Recommended: Mid-range GPU (6-8GB VRAM)
Recommended Models:
  • STT: Whisper Medium, Parakeet v3, Whisper Turbo
  • AI: Ollama 7B-8B models (llama3.2, mistral, gemma2)
Use Case: Fast offline AI with high accuracy

​
Tier 4: Enthusiast

Hardware:
  • 32+ GB RAM
  • 8+ CPU cores
  • High-end GPU (12+ GB VRAM)
Recommended Models:
  • STT: Any model, Whisper Turbo for best accuracy
  • AI: Ollama 13B-30B+ models (llama3.2:13b, mixtral)
Use Case: Maximum quality offline experience
Your performance tier is calculated automatically and displayed in Settings > Offline along with customized model recommendations.

​
Model Recommendations Based on Hardware

​
Suggestion Algorithm

SlasshyWispr recommends models based on:
  1. Total RAM: Must fit model + OS + headroom
  2. CPU cores: More cores = better performance with larger models
  3. GPU presence: Enables acceleration tier recommendations
  4. VRAM: Determines max GPU-accelerated model size

​
Suggested Models Array

The suggestedModels array contains models that should work well on your system, ordered by recommendation priority. Example (16GB RAM, 8 cores, RTX 3060): 
{
  "slasshySuggestionModel": "nvidia/parakeet-tdt-0.6b-v3",
  "suggestedModels": [
    "nvidia/parakeet-tdt-0.6b-v3",
    "openai/whisper-medium",
    "openai/whisper-small",
    "FunAudioLLM/SenseVoiceSmall",
    "UsefulSensors/moonshine-base"
  ]
}

​
Caution Models Array

The cautionModels array lists models that may struggle on your hardware:
  • May exceed available RAM
  • Could cause slow performance
  • Might require swap/paging
Example (8GB RAM system): 
{
  "cautionModels": [
    "openai/whisper-large-v3-turbo",
    "openai/whisper-large-v3"
  ]
}
You can still select caution models, but expect reduced performance. The selectedModelWarning field will explain potential issues.

​
Selected Model Warning

If you choose a model outside recommendations, selectedModelWarning provides specific guidance:
  • “This model may exceed available RAM”
  • “Consider a smaller model for better performance”
  • “GPU acceleration recommended for this model”
  • Empty string if selection is optimal

​
Checking Your Hardware

1

Open Settings

Navigate to Settings > Offline in SlasshyWispr
2

View Hardware Analysis

Your hardware details are displayed:
  • CPU model and core count
  • Total RAM
  • GPU detection status
  • Performance tier
3

Review Recommendations

SlasshyWispr highlights:
  • Top recommended model (green)
  • Other suggested models
  • Models to avoid (red/yellow warnings)
4

Select Optimal Model

Choose from suggested models for best performance

​
Optimizing Performance

​
For Low-Resource Systems

1

Choose Lightweight Models

Use Moonshine or SenseVoice for STT
2

Close Background Apps

Free RAM by closing unnecessary applications
3

Use Hybrid Mode

Local STT + online AI for balanced resource usage
4

Disable GPU Acceleration

If GPU causes issues, force CPU-only mode

​
For High-Performance Systems

1

Enable GPU Acceleration

Ensure NVIDIA drivers and CUDA toolkit are installed
2

Use Larger Models

Whisper Turbo for STT, 13B+ models for Ollama
3

Keep Models Warm

Enable daemon mode to keep models loaded in memory
4

Monitor Latency

Check pipeline latency in Settings > Pipeline

​
Storage Requirements

​
Disk Space Needed

  • Application: ~300 MB
  • STT Models: 58 MB - 1.6 GB per model
  • Ollama Models: 1 GB - 40+ GB per model
  • Voice Models (Piper): ~10-50 MB per voice
  • Cache & Logs: ~100-500 MB
Total Estimate: 5-50+ GB depending on models
Use SSD storage for best model loading performance. HDDs will work but may increase warmup times.

​
Troubleshooting Hardware Issues

​
GPU Not Detected

Issue: NVIDIA GPU not showing in hardware advisor Solutions:
  • Install latest NVIDIA drivers
  • Install CUDA toolkit (11.x or 12.x)
  • Restart SlasshyWispr after driver installation
  • Check GPU visibility: nvidia-smi (Linux/Windows)

​
Insufficient RAM Warnings

Issue: Cannot load model, out of memory errors Solutions:
  • Close other applications
  • Choose a smaller model
  • Increase system swap space (temporary solution)
  • Upgrade RAM for better experience

​
Slow Performance

Issue: Models load but transcription/inference is slow Solutions:
  • Check if swap/paging is active (indicates RAM pressure)
  • Reduce model size
  • Enable GPU acceleration if available
  • Close background processes
  • Check CPU throttling (thermal issues)

​
Model Warmup Failures

Issue: Model fails to warm up or load Solutions:
  • Verify sufficient disk space for model files
  • Check available RAM exceeds model size + 2GB
  • Review logs for specific errors
  • Try re-downloading the model
  • Update to latest SlasshyWispr version
The hardware advisor automatically refreshes recommendations when you install new RAM, upgrade GPU, or change system configuration.

Build docs developers (and LLMs) love

Get started for freeTalk to us