Training model in pySLGR

[vorugant]

Hi, I am trying to use pySLGR to develop a audio to text transcription system. After going through the example programs and documentation, I was wondering does pySLGR provide methods only for feature extractions or does it also provide a model to predict the content present in audio which can be used to convert to text?
@chrismattmann

[JosephPCampbell]

Thank you for your interest in pySLGR. It is intended for automatic speaker, language and gender recognition. Are you trying to convert speech to text? That is a related, but different technology for automatic speech recognition. Fortunately, there is high-quality open-source software for speech-to-text conversion called Kaldi: http://kaldi-asr.org/. Are you working with the amazing Chris Mattmann? We’re #onit! Joe

…

-- Joseph P. Campbell, PhD, FIEEE e-mail: jpc@ll.mit.edu <mailto:jpc@ll.mit.edu> Group Leader, HLT Group 52 office: [REDACTED] Human Language Technology mobile: [REDACTED] MIT Lincoln Laboratory assistant: [REDACTED] 244 Wood St, S2-265 fax: [REDACTED] Lexington, MA 02421-6426 Massachusetts Institute of Technology http://www.ll.mit.edu/mission/cybersec/cybersec-bios/campbell-bio.html From: Hruday Kumar Voruganti [mailto:notifications@github.com] Sent: Sunday, October 15, 2017 7:19 PM To: mitll/pyslgr <pyslgr@noreply.github.com> Cc: Subscribed <subscribed@noreply.github.com> Subject: [mitll/pyslgr] Training model in pySLGR (#1) Hi, I am trying to use pySLGR to develop a audio to text transcription system. After going through the example programs and documentation, I was wondering does pySLGR provide methods only for feature extractions or does it also provide a model to predict the content present in audio which can be used to convert to text? @chrismattmann <https://github.com/chrismattmann> — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1> , or mute the thread <https://github.com/notifications/unsubscribe-auth/AG2OwgsED_IzN1rje_lbCXogqtkfxABDks5sspLVgaJpZM4P58Aa> . <https://github.com/notifications/beacon/AG2Owh2xVfvpcu9OQiRlFtKGSkfIRopiks5sspLVgaJpZM4P58Aa.gif>

[vorugant]

Hi Joe,

Yes I am working on speech to text conversion. So you would suggest to use KALDI over pySLGR for speech to text conversion?
Can I get some more explanation on your comment saying "That is a related, but different technology for automatic speech recognition. " It would be a great help.

Yes I am working with Chris Mattmann.

Thank you
@chrismattmann

[JosephPCampbell]

Hi! Yes, Kaldi is for automatic speech recognition (automatic text-to-speech conversion) and pySLGR is not. pySLGR is for automatic speaker, language and gender recognition (plus speech activity detection). The next questions are what language(s) do you want to recognize speech in, in what domain, and in what conditions (for example, conversational English cellular telephone speech is a common). Please say hello to Chris! Joe On Oct 15, 2017, at 11:50 PM, Hruday Kumar Voruganti <notifications@github.com> wrote: Hi Joe, Yes I am working on speech to text conversion. So you would suggest to use KALDI over pySLGR for speech to text conversion? Can I get some more explanation on your comment saying "That is a related, but different technology for automatic speech recognition. " It would be a great help. Yes I am working with Chris Mattmann. Thank you @chrismattmann — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[vorugant]

Hi Joe,

We actually tried using KALDI before I joined in the project and the results were not very satisfactory.
Thank you for the clarification.
Yes I would say Hi to Chris on behalf of you.

@chrismattmann

[chrismattmann]

@vorugant please re-open this issue. Just saw being flagged in here :)

@JosephPCampbell great to hear from ya. Yes I recommended looking into PySLGR. The activity we are trying to do here is to take many calls from a customer call service center with both customers/clients and agents on the other end (so need to do speaker dilation) and turn audio files into text. We don't have a ton of labeled transcripts for the calls, but we could generate some. We had negative results with KALDI but only I believe b/c we didn't properly train it. Do you have any suggestions?

[JosephPCampbell]

Thanks Chris. This is more than a lunch-break project. Let me see if there's a way for me to have our Kaldi experts help. I'd need to tie it to our sponsored work - hmmm... Another option is JHU's Kaldi team. Joe Joseph P. Campbell, PhD, FIEEE Group Leader, Human Language Technology MIT Lincoln Laboratory Office: [REDACTED] Mobile: [REDACTED] http://www.ll.mit.edu/HLT Sent from my iPhone On Oct 17, 2017, at 2:04 PM, Chris Mattmann <notifications@github.com> wrote: @vorugant please re-open this issue. Just saw being flagged in here :) @JosephPCampbell great to hear from ya. Yes I recommended looking into PySLGR. The activity we are trying to do here is to take many calls from a customer call service center with both customers/clients and agents on the other end (so need to do speaker dilation) and turn audio files into text. We don't have a ton of labeled transcripts for the calls, but we could generate some. We had negative results with KALDI but only I believe b/c we didn't properly train it. Do you have any suggestions? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[chrismattmann]

thanks @JosephPCampbell totally understand. Let us know in the meanwhile we'll keep working on it