Running Whisper Locally for Podcast Subtitles

Situation

We had a 2-hour Nepali podcast with four people speaking. Writing English subtitles manually was taking forever, so we needed a faster way.

Complication

Most subtitle or translation tools don’t support Nepali well. And even if they do, they often can’t handle long audio files or are too expensive to use regularly.

Questions We Asked

• Is there any AI that can help with this?

• How do other podcasters handle it?

• Can we do it cheaply?

• What model do most existing tools use?

Solution

We found OpenAI’s Whisper model, which is built for speech recognition, transcription, and translation.
Since we had a powerful computer, we decided to run it locally instead of using a paid tool.

It worked really well — translating our 2-hour Nepali podcast into English subtitles (.srt file) in about 30–40 minuteswith around 90% accuracy.

How We Did It

1. Install Whisper
First, install Python (version 3.10 or higher) and make sure ffmpeg is installed on your system. Then open your terminal and run:

pip install -U openai-whisper

If you don’t have ffmpeg installed, you can install it with:

brew install ffmpeg

(for Mac) or

sudo apt install ffmpeg

(for Ubuntu/Linux).

2. Prepare the Audio File
Convert your podcast audio into a supported format (like .wav or .mp3) if it isn’t already.
You can use ffmpeg for that:

ffmpeg -i podcast.mp4 podcast.wav

3. Run Whisper for Translation
Once everything is set up, run this command in your terminal:

whisper podcast.wav --model medium --language Nepali --task translate

This tells Whisper to translate Nepali audio into English text using the medium model.
If you want faster results (with slightly lower accuracy), you can use the small model instead:

whisper podcast.wav --model small --language Nepali --task translate

4. Export to Subtitle Format
Whisper automatically creates an .srt subtitle file along with the transcript text file.
You can open and edit it with any text editor if you want to correct spelling or timing issues.

Output files usually include:

podcast.wav.txt      # Full text transcript
podcast.wav.srt      # Subtitle file with timestamps

5. Review & Finalize
We reviewed the output, fixed a few translation issues (mostly names or short phrases), and synced it with the video. The result had about 90% accuracy and saved hours compared to doing it manually.

Outcome

By running Whisper locally, we saved money, kept our data private, and created high-quality subtitles much faster. Now we use the same workflow for all our long-form podcasts and video projects.