Shipping Gemma 4 speech recognition in a Windows .NET desktop app: a 5-variant model-selection tour
Parlotype, a voice-to-text desktop app for Windows built with .NET 10 and Avalonia UI, now supports Gemma 4 as an alternative speech engine alongside Whisper.net. The app captures audio via WASAPI, applies Silero VAD, and injects text into any active application. Google released Gemma 4 in April 2026 with native multimodal audio. The developer evaluated five GGUF variants from the ggml-org repository: E2B and E4B in BF16, Q4_K_M, and Q8_0 formats. Each was benchmarked on the same dataset to determine the optimal combination of accuracy, speed, and disk footprint. The selected default model is shipped with the app, and users can choose between Whisper and Gemma 4 in settings. All processing runs locally, ensuring no audio leaves the machine. The post details the model selection process and provides insights for developers integrating Gemma 4 into .NET applications.
Enables fully offline, low-latency speech recognition in .NET desktop apps using Gemma 4.