We were requested to transcribe several hours of anthropology talks (in Spanish) recorded in mp3 audio files. The audio files had background noise and echo causing speech recognition to fail. We cleaned the noise from the audio files with scripts. The automated cleaning process was effective and efficient.
ejemplo_corto.mp3
Note the noise between the spikes. These areas represent the silence of the speaker.
ejemplo_corto_clean.mp3
Note the lack of noise between the spikes. These areas represent the silence of the speaker.
Speech recognition is a new technology that allows to automatically transcribe audio conversations to text (Wiki: Speech Recognition). There are many technologies that allow speech recognition. While I prefer to use a local engine (nothing is shared) I tried with a few examples using a SpeechRecognition python library (GitHub: PythonPackage-SpeechRecognition). This python library allows running several speech recognition services and libraries which make a comparison easy and efficient. From the options tried, I remark the CMUsphinx (open source and works offline) (Webpage: CMUsphinx) and Google's API (Webpage: Google API for Speech Recognition).
Sadly neither of the two worked as expected due to the audio quality. Therefore, I set myself to clean the audio files from the noise.
After some research (and trial and errors), I found that two tools can help me with this task: ffmpeg (Website: ffmpeg) and sox (Website: sox). There are automatic ways to clean the noise from audio files, but the results were not great. Some extra research lead me to try:
Find a sample of the noise.
Generate a noise profile.
Clean the audios using the noise profile.
The organizers of the conference were quite happy with the results.
[1]: http://www.zoharbabin.com/how-to-do-noise-reduction-using-ffmpeg-and-sox/
[2]: https://en.wikipedia.org/wiki/Speech_recognition
[3]: https://github.com/Uberi/speech_recognition#readme
[4]: https://cmusphinx.github.io/wiki/
[5]: https://cloud.google.com/speech-to-text/
[6]: https://ffmpeg.org/
Library ffmpeg: "A complete, cross-platform solution to record, convert and stream audio and video."
[7]: http://sox.sourceforge.net/
Library sox: "Welcome to the home of SoX, the Swiss Army knife of sound processing programs."
