Dear,
It is less than 10% working for the TensorFlow Lite micro_speech recognition example which is “yes” or “no” classification with SparkFun Edge.
I believe it is because of the quality of Microphone and the ADC.
I’ve captured and measured the performance of the microphone’s PCM data, and I’ve realized that the data was too noisy.
The data were:
-
Too noisy even if at silent environment (see below captured data).
-
Too small the recorded voice level but the base noise was too high.
Pls note that the S/W code used was sparkfun_edge/audio_provider.cc.
Do you know why the quality of Microphone was too bad?
Pls let me know how to resolve this issue as I want to have good recorded voice with very small noise.
Pls note that I got the almost same result with the other Sparkfun EVK board.
Thank you.
//Daum
=========below, captured PCM data at silent environment=========
FF4B,0029,0000,00A7,FFCE,0093,00A9,0066,FFB6,FE41,025E,0022,FFFB,FF9E,FF14,0152,
00FA,FF79,FE12,0071,0216,0036,017D,FFE4,FFA5,0220,01A6,FFEE,FFE0,FF34,FE44,00AE,
FF68,FF99,01EE,01B2,002A,FEAA,0056,0176,00B6,00B2,FFD4,025D,007F,020A,038D,01AA,
0165,00F8,009C,0172,0015,FFE9,FFE1,018E,0195,FF6E,023B,004A,0115,0056,01A8,015C,
0089,0054,FFBD,0099,0038,016B,0139,007C,00DB,0155,01EE,00A5,FFD1,FFE3,0053,0146,
FD51,FE41,FEF6,FE41,FEA4,008F,FFDD,002A,FF85,00FA,00A6,FFD3,FFCC,FDE4,004E,01A2,
FFC8,FED8,FF48,FFC7,00BC,00DD,0098,00D3,FFDA,FF3F,FEE8,010C,0003,FD6F,00FA,0109,
FEA8,02BE,FEB4,023B,FE77,0392,00F9,00AA,0032,FEC3,00A6,FEF8,011A,FEB1,FFC3,0174,
003B,0023,FEA0,0067,0134,00EC,00F0,FF61,012D,FDE4,0149,FFA8,012A,0064,FFF4,FF4E,
0111,0008,FFB3,FFC1,0166,0070,0189,FE77,FD5B,002C,0143,0012,0043,FF36,0245,015A,
0004,01E6,FD80,FEEE,012C,01A8,FFAD,0135,0106,FE6A,013D,001C,FFDB,0014,FC99,FFB0,
FD9D,01DA,FFD4,FF38,FFD1,019F,027D,0152,FC0B,0188,0096,0023,0058,00A3,FE20,012B,
FF80,FF8E,FF96,FF17,0027,013B,FF7A,0109,FED8,0000,011F,0066,FF64,FFFC,01F5,01B6,
0133,FF50,000A,013F,01E4,FFCE,0040,FD70,002B,FFA2,0209,FEF8,011A,0073,FE3E,0376,
==end==