Canary Qwen 2.5B vs Whisper Large V3: Windows setup

This tutorial demonstrates local subtitle generation in SECourses' Whisper TTS Premium with NVIDIA Canary Qwen 2.5B and compares it with Whisper Large V3. On his own tutorial videos, the speaker reports a 5.91% global word error rate for Canary and transcription speeds up to 46 times real time. These results describe his tests, rather than a general performance guarantee. He also notes a few cases where Whisper performed slightly better.

The app accepts audio and video files, processes folders with optional subdirectories, and saves subtitle and transcript formats including VTT, SRT, TXT and TSV. The walkthrough covers chunk length, saved presets, YouTube URLs and microphone transcription. The speaker describes translation quality as limited. Canary lacks word-level timestamps in this demonstration; Whisper remains an option for those timestamps and broader spoken-language support.

The Windows setup uses Python 3.11 and an installer that creates a virtual environment. The speaker lists Git and CUDA requirements, says C++ tools may be unnecessary because of precompiled libraries, and advises avoiding spaces in the installation path. Canary downloads roughly 5 GB on first use. He also covers Linux installation files and updated NVIDIA drivers for CUDA 13 and RTX 5000 series support. For this local AI workflow, he recommends subprocess mode to release RAM and VRAM after transcription.