Help with Transcriber

[Balaji]

I am trying to transcribe text from a wav file of voices of people with speech disability.
The output of (StreamSpeechRecognizerOBJECT.getResult()).getHypothesis() is very much different from the correct word. I am trying to make the recognizer adapt to my test data.

Is there a way that I could look at the phonemes that are recognized by the Sphinx4 engine? I want to see something like this:
K AH M AE N D
for the word COMMAND

I guess the recognizer will first get the phonemes and construct words from them. Am I right in my understanding? Please clarify.

Thank you.

[Nickolay V. Shmyrev]

Is there a way that I could look at the phonemes that are recognized by the Sphinx4 engine?

There is no API for phonemes, you have to modify the source code to return them.

I guess the recognizer will first get the phonemes and construct words from them. Am I right in my understanding? Please clarify.

No, it does not work this way. The recognizer looks for the whole word sequences.

[Balaji]

Thank you Mr. Nickolay.
I have tried using the en_us model as in transcriberdemo.java and the adaptation method as described in CMU tutorial. Is it possible to create the acoustic model afresh from scratch. I have read theories of forward algorithm, forward - backward algorithm / baum-welch algorithm. Could you please give some pointers on this.

Thank you.

[Balaji]

Hello,

In my setup, I transcribe about many .wav files at one shot. I am using code from Transcriber.java as given in the tutorial. I have included the recognizer code inside a loop which reads one .wav file at a time and pass it as a parameter to the recognizer. Something like this:

for (String FileName:FileList) {
    InputStream stream = new FileInputStream(new File(FileName));
    if (stream != null) {
                    // Simple recognition with generic model
                    recognizer.startRecognition(stream);
                    ...
     }
}

I have 2 questions.
1. Can I execute recognizer.startRecognition once outside the loop and pass each wave file as a parameter to the recognizer(using some function I guess)? My idea here is to start the recognizer only once (all the overheads will be done at once), so that the recognizer is continuously listening to some stream. I will send the .wav files one by one to this stream. Any performance improvement possible this way?

1. Part of the output I have given below. In the Timer output, why is the dictionary loaded again and again. The "Load AM" count is only one but, Load Dictionary count is increasing.

15:51:55.510 INFO speedTracker # ----------------------------- Timers-------------------------------
15:51:55.510 INFO speedTracker # Name Count CurTime MinTime MaxTime AvgTime TotTime
15:51:55.510 INFO speedTracker Load Dictionary 765 0.0040s 0.0000s 0.5360s 0.0093s 7.0820s
15:51:55.510 INFO speedTracker ** Load AM 1** 3.5080s 3.5080s 3.5080s 3.5080s 3.5080s

filler dictionary / noisedict is also loaded again and again. Any specific reason why?

In short, I am trying to make my recognizer work faster in decoding many files. By modifying these I feel there may be some speed I can gain. Is it feasible.

Thanks for your help.

Balaji. (I have attached the output trace as a text file).

[Nickolay V. Shmyrev]

Nvidia recently announced audio recognition at speed 3500 times faster than realtime. You should better look on their job:

https://devblogs.nvidia.com/nvidia-accelerates-speech-text-transcription-3500x-kaldi/