A Simple Test of Using WhisperDesktop for Speech-to-Text

TLDR

• WhisperDesktop is an offline speech-to-text tool based on OpenAI Whisper that runs without a Python environment.
• It is recommended to prioritize the ggml-medium.bin model to achieve the best balance between performance and accuracy.
• For users with a dedicated graphics card, processing 5 minutes of audio with ggml-medium.bin takes only about 11 seconds.
• For users without a dedicated graphics card, ggml-small.bin is recommended as the baseline for daily use, as ggml-tiny.bin has insufficient accuracy.
• The developer has stopped updating WhisperDesktop; it is recommended to switch to Subtitle Edit with Faster-Whisper integration for better performance and maintenance support.

WARNING

The WhisperDesktop developer has not updated the project for a long time. It is currently recommended to switch to Subtitle Edit with Faster-Whisper integration, which is more actively maintained and faster. For details, please refer to: Using Subtitle Edit with Faster-Whisper for Local Speech-to-Text.

Download and Installation

WhisperDesktop is a lightweight offline tool that does not require a Python environment.

1. Go to the WhisperDesktop GitHub Releases page to download the latest version.
2. After unzipping, you will find WhisperDesktop.exe (the executable) and Whisper.dll (the library).

Model Selection and Specifications

Models must be downloaded from Huggingface Whisper. The model size directly affects VRAM requirements and processing speed:

Size	Parameter Count	Required VRAM	Relative Speed
tiny	39 M	~1 GB	~32x
base	74 M	~1 GB	~16x
small	244 M	~2 GB	~6x
medium	769 M	~5 GB	~2x
large	1550 M	~10 GB	1x

Usage

When might you encounter configuration issues: If the software cannot automatically detect your hardware, you need to adjust the parameters manually.

1. Run WhisperDesktop.exe and specify the model path.
2. Set Model Implementation to GPU. If the graphics card cannot be detected, you can click advanced... to configure it manually.
3. Select the target language in Language.
4. Output Format supports various formats such as .txt, .srt, and .vtt.
5. Check Place that file to the input folder to save the output file directly in the input file directory.

Performance Test Results

When might you encounter performance bottlenecks: Using an overly large model (such as large) on certain hardware may lead to processing failures or empty output.

• Dedicated Graphics Card (RTX 4070 Ti Super):
  • ggml-medium.bin: Processing 5 minutes and 16 seconds of audio takes only 11 seconds.
  • ggml-large-v3.bin: Processing time takes up to 22 minutes, with a risk of conversion failure.
• Integrated Graphics (i7-12700H):
  • ggml-tiny.bin: 41 seconds.
  • ggml-small.bin: 4 minutes and 19 seconds.
  • ggml-medium.bin: 13 minutes and 5 seconds.

Conclusion and Recommendations

• Dedicated Graphics Card Users: It is recommended to use ggml-medium.bin consistently, as it provides the most stable performance and accuracy.
• Integrated Graphics or Older Graphics Card Users:
  • For daily transcription, ggml-small.bin is recommended as the minimum threshold for accuracy.
  • If high-precision content is required, you can choose ggml-medium.bin and allow for a longer processing time.

Changelog

1. 2025-03-24

   • Initial document created.
2. 2026-01-31

   • Added recommendation link to the new Faster-Whisper solution.