update csv output format to match OpenAI's Whisper dataframe output

[hykelvinlee42]

Currently, the csv output file does not have column field names. And when the file is read by data analysis tool (pandas e.g.), it will raise parsing and tokenizing errors.

import pandas as pd

pd.read_csv(output.csv, names=columns)
# pandas.errors.ParserError: Error tokenizing data

This can be resolved by setting delimiter to a whitespace character in the read function. However, this would cause the start timestamp and end timestamp to be read as string with commas by default.

pd.read_csv(output.csv, names=columns, delimiter=" ")
"""
      start      end                                               text
0        0,   10320,   We're not actually looking at on the screen, ...
1    10320,   18100,        And you can see that recording has started?
2    18100,   23600,                       Recording and transcription.
"""

This change allows the result csv file to be read successfully by default with the start and end timestamps as int64 type.

[jordibruin]

I'm curious, with this format, how the csv will not mess up when there's commas in the text? Won't that trigger a 'next item' in the csv parsing?

[hykelvinlee42]

@jordibruin the text is already encapsulated in double quotes ("). So pandas, or any csv parser tools that I know of, will parse everything within the double quotes as string, even commas.

whisper.cpp/examples/main/main.cpp

Line 362 in 86cddf8

fout << 10 * t0 << "," << 10 * t1 << ",\"" << text << "\"\n";