Models Library & Speech Engine

Dicty features a modular AI architecture supporting on-device hardware-accelerated models and ultra-accurate cloud intelligence.

---

Local Speech Models (ASR)#

Local models run 100% offline on your Mac using whisper.cpp optimized with Apple Metal GPU shaders.

Model Specifications & Quality Inspection#

Click the info badge next to any installed speech model to inspect its technical profile, installed quantization tier, and relative benchmarking stats:

• Status Indicator: Shows Ready • Offline when the weights are validated and available for instant offline transcription.
• Installed Version: Displays active disk footprint (e.g., 140 MB for Base Q5).
• Benchmark Gauges: Visual meters for real-time inference speed (e.g., 80%) and acoustic accuracy (e.g., 60%).

Specialized Regional Models#

• BELLE (Chinese - 3.1 GB): Fine-tuned for Mandarin and Chinese dialects.
• Large V2 (French - 3.1 GB): Trained on thousands of hours of French audio to resolve silent letter ambiguities.
• Large V3 Turbo (Japanese - 1.6 GB): Fine-tuned to eliminate Japanese hallucinations.

---

Understanding Quantization & Download Tiers#

When downloading local models from the library, Dicty provides three precision profiles tailored for different hardware setups:

1. Optimized Studio (Q5) — Recommended:
   • Retains ~99% of full 16-bit acoustic accuracy while cutting memory and disk requirements by 50% (e.g. 3.1 GB for Large V1/V3).
   • Provides optimal performance on Apple Silicon unified memory (M1 or later).
2. Uncompressed (F16) — Maximum Precision:
   • 16-bit uncompressed floating-point weights (~10.2 GB).
   • Best for studio environments, heavy background noise, or subtle phoneme distinctions.
3. Ultra Light (Q4) — Minimal Footprint:
   • 4-bit integer weights (~2.3 GB).
   • Blazing-fast inference speed and minimal battery usage for laptop multitasking.

---

LLM Text Refinement & Advanced Model Settings#

Dicty pairs raw speech recognition with local and cloud Large Language Models to format book prose, remove verbal disfluencies, and draft structured notes:

LLM Text Refinement Models#

• Local GGUF Models: Download and run lightweight instruction models locally (e.g. Qwen 2.5 1.5B for fast offline punctuation, Llama 3.2 3B for comprehensive editing).
• Cloud LLMs: Connect to top-tier reasoning engines (DeepSeek V3, GPT-4o Mini, Qwen 2.5 72B, Claude 3.5 Sonnet) with one click.

Advanced Model Settings#