JhonatanAiT14/dictate.sh
🎤 Transcribe speech with low-latency on Apple Silicon using dictate.sh; offering real-time ASR and intent analysis in a simple, offline setup.
What's novel
🎤 Transcribe speech with low-latency on Apple Silicon using dictate.sh; offering real-time ASR and intent analysis in a simple, offline setup.
Code Analysis
10 files read · 5 roundsA single-file real-time speech-to-text application for Apple Silicon that implements a custom MLX-based audio encoder (convolutions + transformer layers) and text decoder, with WebRTC VAD for turn detection, rolling-window ASR for low latency, and optional LLM-based intent analysis — all running loc
Strengths
Genuine implementation depth with a full custom neural network architecture (audio encoder with 3D convolutions, sinusoidal positional embeddings, 24 transformer layers, block-wise attention masking) and a well-designed real-time audio pipeline with VAD, rolling windows, and GPU serialization. The code is well-documented with clear design notes, sensible defaults, and thoughtful separation of stdout (transcripts) from stderr (logging).
Weaknesses
Zero tests for a complex real-time audio/ML pipeline is a significant gap. The entire application is a single 63KB file with no modular separation, making it hard to test, extend, or maintain. The README is a misleading fork that replaces the project name with a raw GitHub zip URL, and the docs/stt.py is a redundant duplicate.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
17
Contributors
2
Files
9
Active weeks
4
Repository
Language
Python
Stars
1
Forks
0
License
MIT