"Can You Hear Me Now?"

GoPiGo
,
1

Carl began life with a microphone I had tested successfully with the PocketSphinx speech recognition engine back in 2016.

This is the Kinobo Mini Akira:

Carl_Oct2019
Carl_Oct20193024×4032 2.01 MB

I always had suspicions that the mic might be a bit weak in sensitivity

  • Kinobo Mini Akira: Sensitivity -47dB +/-4dB
     

even though it performed “ok” for far-field speech recognition on Carl located on the floor.

 

Far-Field Microphones

Usually far-field recognition requires special microphone arrays such as included in the Google AIY Voice Kit or the ReSpeaker system of RPi hats and mic arrays.
 

Can a normal Mic Compete?

Acting on my desire to give Carl the best “ear(s)” consistent with his crowded real estate and already capped off GPIO connector, I purchased the Movo MA200GY Omni directional Electret Condenser microphone with USB audio interface, that claims considerably higher sensitivity of:

  • Movo MA200GY Omni directional Electret Condenser
    Sensitivity: -30dB +/-3dB / 1kHz 0dB=1V/Pa, S/N: 74dB SPL,
     

Carl_Mic_Movo_MA200GY
Carl_Mic_Movo_MA200GY3024×4032 1.58 MB

 

Samples:

I tried to figure out how to post samples here, but cannot.

Speech Level, Background Level, and S/N

Old Kinobo Mini Akira:

SN_s1-m1-hey_carl_18
SN_s1-m1-hey_carl_18855×605 108 KB

  • RMS Volume: -26
  • Background: -55
  • S/N 30 dB
     

New Movo MA200GY:

SN_s1-m2-hey_carl_01
SN_s1-m2-hey_carl_01841×615 69.1 KB

  • RMS Volume: -18
  • Background: -41
  • S/N: 23dB

Higher volume but lower signal to noise? How will this affect recognition? TBD

Frequency Response (to speech, not white source)

Old Kinobo Mini Akira:
Spectrum_s1-m1_hey_carl_18

 

New Movo MA200GY:

Spectrum_s1-m2-hey_carl_10

  • Seems to be much flatter response.

 

How will this affect recognition? TBD

 
 

To be complete, I also tested:

image
image679×833 53.3 KB

  • Kinobo - USB 2.0 Mini Microphone Makio Mic
    Sensitivity:-67 dBV/pBar,-47dBV/Pascal± 4dB

 

SN_s1-m3-hey_carl_10
SN_s1-m3-hey_carl_101574×1230 161 KB

  • Very Low Sensitivity
     

Spectrum_s1-m3-hey_carl-01
Spectrum_s1-m3-hey_carl-011094×950 88.1 KB

  • Not great but not necessarily bad response

 

Summary of investigation

Testing with the Nyumaya Hotword engine and the Vosk-API engine with both small language model and word list modes, show both the first Kinobo Mini Akira and the second Movo MA200GY microphones allow far-field wake up, and far-field command phrase recognition.

The Movo mic eliminates the “cord mess” from Carl’s side, and seems to work well even when I speak to Carl from another room. I’m going to keep using the Movo.

Hey Carl, Can You Hear Me Now?

2 Likes
2

Yea, but “Da’ Bitch Part” is they all absolutely vacuum above about 800 Hz, so you really loose a lot of the high frequency details - which is where a lot of the distinguishing sound data is.

Even with the Movo MA200GY you’ve lost 30 db at about 2 kHz, (with a peak that’s only 25 db down at 2.5 kHz), and by the time you get to 5 kHz, you’re in the grass.

How are you measuring this?  That can make huge amounts of difference.

From looking at the data, saying the Movo is “flatter” is like saying that being hit with a 20 lb sledge hammer wrapped in a handkerchief is less painful than being hit by the sledge without the handkerchief.  No matter which sledge you get hit with, your head is going to be in Louisiana when it finally lands!

1 Like