When a non-english dub exists of an anime, it can still be transcribed with the English model.

[jinocenc]

Example: If there is only a Japanese audio file then the Japanese audio file is transcribed as English and saved to the eSRT file.

The issue could be solved by checking for multiple streams and by checking for English streams.

This code block expects an error thrown from FFMPEG

whisper_grpc/whisper_grpc/transcribe.py

Lines 20 to 28 in e5a1690

	try:
	output_file = f'{tmpdir}/{file}.wav'
	transcribe_audio_file(file, output_file)
	result = whisper.transcribe(
	model, output_file, beam_size=5, best_of=5, verbose=logger.getEffectiveLevel() <= logging.DEBUG,
	language="en")
	except ffmpeg.Error as e:
	logger.error(f'No English audio track found in {file}, error: {e.stderr}')
	raise ValueError(f'No English audio track found in {file}')

The error should be thrown here

whisper_grpc/whisper_grpc/transcribe.py

Lines 35 to 43 in e5a1690

	def transcribe_audio_file(input_file, output_file):
	(
	ffmpeg
	.input(input_file)
	.audio(metadata='language=eng')
	.output(output_file, output_format='mp3')
	.overwrite_output()
	.run()
	)

--Except, I'm pretty sure this library doesn't support English-stream identification

whisper_grpc/Pipfile

Line 11 in e5a1690

python-ffmpeg = "*"

suggested actions:

• migrate to working(?) library
• properly probe file for English audio streams using language meta-data
• make sure exception is properly thrown

[Juxsta]

Actually not using python-ffmpeg in that code block. I had an issue with figuring out how to find the English track, but it worked correctly with ffmpeg directly. I believe the language=eng filter should grab the first English audio track, and it'll throw an error otherwise. There could be an issue if the audio track is incorrectly tagged