General Discussion

Nested
Bluetooth device as audio in/out
User: colbec
Date: 10/10/2009 6:58 am
Views: 7480
Rating: 9

I've been experimenting with a bluetooth dongle and a Jabra BT2040 as an audio input and output device. The idea is a handsfree communication with a Julius SRE in place of the wired USB headset I normally use. (Using Opensuse 11.1 + Blueman)

However so far the bluetooth interface (and I admit I don't have much experience with it yet) seems to be pretty clunky, each time a prompt is recorded there is a delay of several seconds while the communication is set up, pops and whistles etc which have to be edited out of the sample. It all makes for a lot of work compared to a wired headset.

Maybe I have just chosen the wrong hardware. Has anyone else had luck with bluetooth in this context?

--- (Edited on 10/10/2009 6:58 am [GMT-0500] by colbec) ---

Re: Bluetooth device as audio in/out
User: Robin
Date: 10/13/2009 4:20 pm
Views: 192
Rating: 9

>Maybe I have just chosen the wrong hardware.

 

No experience myself, but that's quite likely, I read it has an auto sleep mode to save battery power. Perhaps a rechargeable always on bluetooth headset is a better option.

--- (Edited on 10/13/2009 4:20 pm [GMT-0500] by Robin) ---

Re: Bluetooth device as audio in/out
User: colbec
Date: 11/14/2009 5:47 am
Views: 114
Rating: 12

Okay guys there is some progress to report here. I installed the new Opensuse 11.2 yesterday and using alsa/pulseaudio with instructions from their website I now have command line interface control which uses aplay and arecord to use the Jabra BT2040 but not my Jabra VBT185Z. This is a big step forward since I can now use the 2040 to record prompts and get feedback. The sound is a bit hummy and there is a slight echo but not enough to interfere with a normal conversation. I will post later with results using the headset in this manner in a Julius/HTK/Voxforge interactive context.

I looked into submitting audio to Voxforge using this headset (as an alternate type of input device, broadening the variety of devices) but using the Voxforge Java web recording interface does not seem to recognize the pulseadio control. Anyone have any experience in directing firefox to use PA?

--- (Edited on 11/14/2009 5:47 am [GMT-0600] by colbec) ---

Re: Bluetooth device as audio in/out
User: kmaclean
Date: 11/16/2009 12:08 pm
Views: 101
Rating: 8

Hi colbec,

>but using the Voxforge Java web recording interface does not seem to

>recognize the pulseadio control. Anyone have any experience in

>directing firefox to use PA?

This might actually be a problem with the Java applet...  It seems simialr to the problem with USB microphones as described in the Java Applet Troubleshooting Guide:

USB Microphones

The Speech Submission applet assumes that you will connect to your default audio source.  If your computer has an analog microphone jack (on your motherboard or on an audio card) it assumes that this is your default audio source, even if you plug in a USB microphone (it basically ignores the audio from your USB microphone).  The Speech Submission app currently has no way to change the audio source (yet..).

Possible approaches to correct this:

  • Disable your motherboard's microphone jack (or the audio card itself) in your system bios (this changed on the menu you can select when your computer first starts up);
  • Remove your audio card; or
  • Modify the sound.properties file in the JRE\lib directory of your JDK installation.

Ken

--- (Edited on 11/16/2009 1:08 pm [GMT-0500] by kmaclean) ---

Re: Bluetooth device as audio in/out
User: colbec
Date: 11/22/2009 8:51 am
Views: 117
Rating: 10

Ken, as an update, I have pretty much given up on getting my bluetooth combo to work on my main working machine which is Opensuse 11. I was hoping that the recent publication of 11.2 would iron out some issues but while there have been major improvements to Kbluetooth4 which now reliably detects both dongle and headsets and apparently allows them to communicate, there is still a fundamental issue at the bluetoothd level which prevents my particular combo of dongle and headsets from cooperating under linux, specifically negotiating an SCO channel. Difficulties reverse engineering the successful Windows comms make it unlikely that this specific hardware will be reliable in my Linux in the near future. Among my options is to just go ahead and buy a number of different dongles and headsets until I stumble on one that works. However this will have to be done gradually over time due to budget constraints. In the meantime the hardware combo works fine with provided software Bluesoleil on Windows. I have found that sound quality with the Jabra VBT185Z headset is much superior to the BT2040 with the same dongle.

Voice submission using the Java applet is fine under windows with bluetooth, using the bluetooth SCO device. For the time being,  my experiments with bluetooth as input/output as SRE interface will use a Windows based recording mechanism.

--- (Edited on 11/22/2009 8:51 am [GMT-0600] by colbec) ---

Re: Bluetooth device as audio in/out
User: colbec
Date: 11/30/2009 7:29 am
Views: 133
Rating: 10

Yet another step: I found a combination of dongle and headset that works for me on Opensuse 11.2 - a Belkin dongle and Plantronics Explorer 360. These two can talk to each other pretty reliably, YMMV.

I completed a project with 352 prompts using a number of one and two word grammar combinations to see what the results are and this is my finding:

====================== HTK Results Analysis =======================
  Date: Mon Nov 30 07:51:34 2009
  Ref : testref.mlf
  Rec : recout.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=57.39 [H=202, S=150, N=352]
WORD: %Corr=71.51, Acc=69.06 [H=379, D=0, S=151, I=13, N=530]
===================================================================

This is horrible and quite unusable in an SRE context, quite unlike my experience with wired headsets. Reviewing the recordings finds there is a bit of hiss, but otherwise clean. Again, it might be a matter of quality of equipment, there are other more expensive dongles and headsets.

--- (Edited on 11/30/2009 7:29 am [GMT-0600] by colbec) ---

Re: Bluetooth device as audio in/out
User: Robin
Date: 12/1/2009 2:44 pm
Views: 78
Rating: 9

I'm not sure, but that might be due to the fact that Bluetooth uses lossless compression. There are some headsets that are approved by/endorsed by some providers who offer speech recognition software. It might be that those headsets have better microphones, but I'm guessing that there are better quality is mostly due to better compression algorithms (or perhaps even lossless compression?). However I'm speculating, since I'm far from an expert in that area.

Nonetheless, I think it is definitely interesting to look at it. I can imagine that there are many interesting applications if this could be done relatively accurate in the future. Never experimented with Bluetooth and Linux so far, so unfortunately I cannot offer any good advice.

--- (Edited on 12/1/2009 2:44 pm [GMT-0600] by Robin) ---

Re: Bluetooth device as audio in/out
User: colbec
Date: 12/4/2009 7:58 am
Views: 149
Rating: 9

Robin:
I looked closely at the recordings for the poor bluetooth results and found a couple of issues
First the volume is low. Viewed in Audacity the peaks hardly reached one third of the waveform. I don't know how much this contributes to the problems. It is also difficult to do anything about this issue. Speaking directly into the mike off-ear results in better volume but occasional severe distortion.
Second many of the prompts were choppy. Choppy in the sense that the waveform looked ok but in fact there were embedded silences. These may have occurred because of timeslicing of resources.

I made some adjustments, changed the recording process, and re-recorded with better results. Here is a set of tests with a Logitech Clearchat Pro wired headset as a baseline followed by 3 tests with various headsets. I did my best to ensure that the only variable was the headset used and of course the prompts had to be re-recorded each time. The results are much better, and probably with a more expensive headset might be quite good.

==============

With Logitech wired clearchat pro

====================== HTK Results Analysis =======================
  Date: Thu Dec  3 15:39:43 2009
  Ref : testref.mlf
  Rec : recout.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=92.61 [H=326, S=26, N=352]
WORD: %Corr=95.09, Acc=94.91 [H=504, D=0, S=26, I=1, N=530]
===================================================================

With Jabra BT2040

====================== HTK Results Analysis =======================
  Date: Thu Dec  3 15:13:05 2009
  Ref : testref.mlf
  Rec : recout.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=87.22 [H=307, S=45, N=352]
WORD: %Corr=91.13, Acc=90.94 [H=483, D=1, S=46, I=1, N=530]
===================================================================

With Jabra VBT185Z

====================== HTK Results Analysis =======================
  Date: Fri Dec  4 08:33:28 2009
  Ref : testref.mlf
  Rec : recout.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=88.64 [H=312, S=40, N=352]
WORD: %Corr=92.45, Acc=91.89 [H=490, D=0, S=40, I=3, N=530]
===================================================================

With Plantronics Explorer 360

====================== HTK Results Analysis =======================
  Date: Fri Dec  4 07:22:38 2009
  Ref : testref.mlf
  Rec : recout.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=86.08 [H=303, S=49, N=352]
WORD: %Corr=90.57, Acc=90.57 [H=480, D=0, S=50, I=0, N=530]
===================================================================

--- (Edited on 12/4/2009 7:58 am [GMT-0600] by colbec) ---

Re: Bluetooth device as audio in/out
User: Robin
Date: 12/6/2009 3:37 am
Views: 84
Rating: 9

Hi colbec,

Those results look promising indeed. I guess the physical design of most Bluetooth headsets (with the microphone ending closer to the ear than to the mouth) doesn't really help. However, since they're almost all constructed similarly, your training speech will need to reflect that I guess.

It's good to see that it is still possible to get half decent recognition results. Perhaps I should also try and record some Bluetooth headset speech... keep us posted!

--- (Edited on 12/6/2009 3:37 am [GMT-0600] by Robin) ---

Re: Bluetooth device as audio in/out
User: colbec
Date: 12/6/2009 9:50 am
Views: 2228
Rating: 13

Robin, here is a test with a Logitech clearchat pro wireless.

With Logitech Wireless Clearchat Pro

====================== HTK Results Analysis =======================
  Date: Sun Dec  6 10:35:47 2009
  Ref : testref.mlf
  Rec : recout.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=85.80 [H=302, S=50, N=352]
WORD: %Corr=90.57, Acc=90.57 [H=480, D=0, S=50, I=0, N=530]
===================================================================

This is an interesting result. It is contrary to my expectation - I was thinking that with a boom mike and a dedicated dongle-headset hardware pair that the result just had to be better than the bluetooth. I was wrong. Without going into a statistical analysis it seems that all the wireless solutions are a magnitude worse than the wired solutions. Still probably workable with an appropriate grammar and diligent choice of words.


My next step is to look carefully at the recordings from each of the setups and see what leaps out.

--- (Edited on 12/6/2009 9:50 am [GMT-0600] by colbec) ---

PreviousNext