ericwang2006/whisper-webui repository overview

⁠Whisper WebUI

A Docker-based WebUI for OpenAI's Whisper speech recognition model with support for both standard Whisper and Faster-Whisper implementations.

The source code is available at https://github.com/ericwang2006/whisper-webui⁠.

⁠Usage

⁠Using Standard Whisper

1. Start WebUI with GPU Support

sudo docker run -d --gpus=all -p 7860:7860 ericwang2006/whisper-webui

1. CLI Usage

sudo docker run --rm \
--gpus=all \
-v ./.cache/whisper:/root/.cache/whisper \
-v ./.cache/huggingface:/root/.cache/huggingface \
-v ./:/app/data \
ericwang2006/whisper-webui \
python3 cli.py --model medium --auto_parallel True \
--vad silero-vad-skip-gaps \
--vad_max_merge_size 8.0 \
--vad_merge_window 1.0 \
--vad_padding 0.3 \
--language Chinese \
--fp16 True \
--output_dir /app/data /app/data/example.mp4

⁠Using Faster-Whisper

1. Start WebUI with GPU Support

sudo docker run -d --gpus=all -p 7860:7860 ericwang2006/whisper-webui:faster-whisper

1. CLI Usage

sudo docker run --rm \
--gpus=all \
-v ./.cache/whisper:/root/.cache/whisper \
-v ./.cache/huggingface:/root/.cache/huggingface \
-v ./:/app/data \
ericwang2006/whisper-webui:faster-whisper \
python3 cli.py --model large --auto_parallel True \
--vad silero-vad-skip-gaps \
--vad_max_merge_size 8.0 \
--vad_merge_window 1.0 \
--vad_padding 0.3 \
--language Chinese \
--whisper_implementation faster-whisper \
--compute_type float16 \
--output_dir /app/data /app/data/example.mp4

⁠Cache Directories

The following directories are used for caching:

• /root/.cache/whisper - Whisper model cache directory
• /root/.cache/huggingface - Faster-Whisper model cache directory

These directories are mounted as volumes to persist downloaded models between container runs.

⁠Additional Notes for Command Line Usage

• To run the CLI in CPU mode, remove the --gpus=all parameter from the command.
• The --language parameter can be used to specify the language of the audio input, which helps improve transcription accuracy.

⁠Key Parameter Explanations

1. --vad_max_merge_size – This is the most critical parameter
   • Setting it between 10.0-15.0 seconds limits the maximum length of a single subtitle segment.
   • If you want shorter subtitles, set it to 8.0 seconds or less.
2. --vad_merge_window – Controls merge logic
   • Setting it to 1.5-2.0 seconds helps reduce over-merging of adjacent voice segments.
   • Smaller values produce more but shorter subtitle segments.
3. --vad silero-vad-skip-gaps – Algorithm choice
   • This variant skips silent gaps compared to the basic silero-vad.
   • It helps split subtitles naturally at speech pauses.
4. --vad_padding – Fine-tunes segment boundaries
   • Setting it from 0.3-0.5 seconds adds a small padding around each segment.
   • Helps avoid cutting off the beginning or end of words.

⁠Adjusting Parameters Based on Content Type

• Dialogue/Interview: --vad_max_merge_size 12.0 --vad_merge_window 2.0
• Speech/Lecture: --vad_max_merge_size 20.0 --vad_merge_window 3.0
• Fast Conversation: --vad_max_merge_size 8.0 --vad_merge_window 1.0

⁠Features

• GPU acceleration support
• Voice Activity Detection (VAD) with configurable parameters
• Support for multiple languages
• Both standard Whisper and Faster-Whisper implementations
• Web UI and CLI interfaces
• Configurable model sizes and precision settings