auto-transcribe

auto-transcribe is a small project built on top of vosk-api to automatically transcribe interviews.

auto-transcribe produces srt files. srt files are human-readable and compatible with many tools which work with audio/video.

I needed a tool that does not require uploading the interview data. Strict data protection guidelines in social sciences usually do now allow uploading participant data to third-party servers if they are not fully anonymized. For this reason, interviews in social sciences are still commonly transcribed manually. This process is extremely expensive and can at least be supported by automatic transcription.

Features

• Transcription of audio data in many languages
• Speaker diarization (works okayish but will need manual processing)
• Multiprocessing

Installation

First install dependencies using poetry

poetry install

Then download the models you would like to use. Checkout the models folder for more info.

Usage

usage: transcribe.py [-h] [--language-model LANGUAGE_MODEL] [--punctuation-model PUNCTUATION_MODEL] [--speakers SPEAKERS] [--speech-enhancement-on] [--ffmpeg-strategy {MERGE,LEFT,RIGHT}] [--log-level LOG_LEVEL] [--threads THREADS]
                     audio [audio ...]

Transcribe interviews.

positional arguments:
  audio                 The audio files to transcribe with the given settings

optional arguments:
  -h, --help            show this help message and exit
  --language-model LANGUAGE_MODEL
                        The path to the vosk language model
  --punctuation-model PUNCTUATION_MODEL
                        The path to the punctuation model
  --speakers SPEAKERS   Number of speakers during the interview
  --speech-enhancement-on
                        Use speechbrain's sppech enhancement. It usually does NOT enhance ASR accuracy
  --ffmpeg-strategy {MERGE,LEFT,RIGHT}
                        Strategy to use for creating a mono file using ffmpeg. If you use stereo microphones, you can select from 'MERGE', 'LEFT' and 'RIGHT'. Using the appropriate channel can improve accuracy.
  --log-level LOG_LEVEL
                        Set the loglevel
  --threads THREADS     Set the number of threads allowed to use for processing multiple audio files

Conversion using ffmpeg

Conversion is now optionla. However, you can convert your interview files to 16 bit 16kHz PCM mono wave

ffmpeg -i in.mp3 -acodec pcm_s16le -ac 1 -ar 16000 out.wav

Occasionally, minor preprocessing (denoising and compressing) improves recognition

ffmpeg -i in.wav -af  "highpass=f=60,lowpass=f=10000,acompressor=threshold=-20dB:ratio=15:attack=0.01:release=0.01,volume=+15dB,afftdn=nf=-20" -acodec pcm_s16le -ac 1 -ar 16000 out.wav

Finally, start recognition.

poetry run python src/transcribe.py --language-model models/vosk/vosk-model-de-0.21 --punctuation-model models/repunc/vosk-recasepunc-de-0.21/checkpoint in.wav

You can adjust the number of expected speakers by adding --speakers <n> to the call. The default is 2.

Limitations

• Make sure that the audio quality is really good. Poor audio quality directly influences the transcipt's accuracy. Poor quality audio will result in non-sense transcripts. Pre-processing of the audio files is certainly possible (and occasionally improves recognition) but needs manual settings.
• The transcript accuracy depends a lot on the language model. For the current German vosk language model the word-error-rate on their podcast testing data is about 20%. Expect to edit the result!