Whisper

Set Up Environment and Whisper

Install Python and the Whisper library (or the optimized faster-whisper version) on your machine or server.

Run `pip install openai-whisper` or `pip install faster-whisper` for faster processing on CPU.

If you have a GPU, additionally install the CUDA-supported version of PyTorch to significantly speed up processing compared to CPU.

Choose Appropriate Model Size

Consider the trade-off between speed and accuracy to select a model size suitable for the existing infrastructure.

Use the 'medium' or 'large-v3' model for Vietnamese as it has significantly higher accuracy than 'base' or 'small'.

'large-v3' model provides the best accuracy for Vietnamese but requires at least 10GB of VRAM to run smoothly.

Run Transcription on Audio File

Run the transcription command with the audio input file, clearly specifying the language as Vietnamese.

Run `whisper file-ghi-am.mp3 --model large-v3 --language vi --output_format srt`.

Clearly specifying `--language vi` helps to increase speed and accuracy, avoiding Whisper from misguessing the language.

Post-Process Punctuation and Formatting

Review the resulting text to correct punctuation and format it for better readability.

Open the result file (.srt or .txt), quickly review the segments for any incorrect punctuation or misrecognized industry terms.

For technical terms that are often misrecognized repeatedly, create a find-replace table to handle bulk processing with a script instead of editing each file manually.

Integrate into Batch Processing Workflow

Write a script to automatically process multiple audio files simultaneously for long-term use in internal workflows.

Write a Python loop to traverse the directory containing audio files, call Whisper for each file, and save results into the corresponding output directory.

Run batch processing overnight or during off-peak hours if processing on a shared server with other tasks.