How to improve cmuSphinx's accuracy? - c

Question

I want to use pocketShpinx to do some speech-to-text word. I have install sphinxbase and pocketSphinx. And download the acoustic model/langauge model/dictionary. Then I test the example code just like follows:
#include <pocketsphinx/pocketsphinx.h>
#include <stdio.h>
#include <stdlib.h>
#include "debug.h"
int main(int argc, char *argv[])
{
ps_decoder_t *ps;
cmd_ln_t *config;
FILE *fh;
int rv;
char const *hyp, *uttid;
int32 score;
config = cmd_ln_init(NULL, ps_args(), TRUE,
"-hmm", "/home/madper/speech/hub4opensrc.cd_continuous_8gau",
"-lm", "/home/madper/speech/language_model.arpaformat.DMP",
"-dict", "/home/madper/speech/cmudict/cmudict/sphinxdict/cmudict_SPHINX_40",
NULL);
if (config == NULL)
{
DBG (("cmd_ln_init() failed.\n"));
exit(1);
}
if ((ps = ps_init (config)) == NULL) /* init decoder */
{
DBG (("ps_init() failed.\n"));
exit(1 );
}
if ((fh = fopen("test.raw", "rb")) == NULL) /* open raw file */
{
DBG (("fopen() failed.\n"));
exit (1);
}
if ((rv = ps_decode_raw (ps, fh, "test", -1)) < 0 )
{
DBG (("ps_decode_raw() error!\n"));
exit (1);
}
if ((hyp = ps_get_hyp(ps, &score, &uttid)) == NULL)
{
DBG (("ps_get_hyp() failed!\n"));
exit (1);
}
printf ("Recognized: %s\n", hyp); /* this is what you say */
fclose(fh);
ps_free(ps);
return 0;
}
DBG is just a macro to print error message if defined DEBUG.
Then I write some code to record from mic use alsa. Like follows:
#define ALSA_PCM_NEW_HW_PARAMS_API
#include <alsa/asoundlib.h>
int main() {
long loops;
int rc;
int size;
snd_pcm_t *handle;
snd_pcm_hw_params_t *params;
unsigned int val;
int dir;
snd_pcm_uframes_t frames;
char *buffer;
/* Open PCM device for recording (capture). */
rc = snd_pcm_open(&handle, "default",
SND_PCM_STREAM_CAPTURE, 0);
if (rc < 0) {
fprintf(stderr,
"unable to open pcm device: %s\n",
snd_strerror(rc));
exit(1);
}
/* Allocate a hardware parameters object. */
snd_pcm_hw_params_alloca(&params);
/* Fill it in with default values. */
snd_pcm_hw_params_any(handle, params);
/* Set the desired hardware parameters. */
/* Interleaved mode */
snd_pcm_hw_params_set_access(handle, params,
SND_PCM_ACCESS_RW_INTERLEAVED);
/* Signed 16-bit little-endian format */
snd_pcm_hw_params_set_format(handle, params,
SND_PCM_FORMAT_S16_LE);
/* Two channels (stereo) */
snd_pcm_hw_params_set_channels(handle, params, 1);
/* 44100 bits/second sampling rate (CD quality) */
val = 16000;
snd_pcm_hw_params_set_rate_near(handle, params,
&val, &dir);
/* Set period size to 32 frames. */
frames = 16;
snd_pcm_hw_params_set_period_size_near(handle,
params, &frames, &dir);
/* Write the parameters to the driver */
rc = snd_pcm_hw_params(handle, params);
if (rc < 0) {
fprintf(stderr,
"unable to set hw parameters: %s\n",
snd_strerror(rc));
exit(1);
}
/* Use a buffer large enough to hold one period */
snd_pcm_hw_params_get_period_size(params,
&frames, &dir);
size = frames * 2; /* 2 bytes/sample, 2 channels */
buffer = (char *) malloc(size);
/* We want to loop for 5 seconds */
snd_pcm_hw_params_get_period_time(params,
&val, &dir);
loops = 2000000 / val;
while (loops > 0) {
loops--;
rc = snd_pcm_readi(handle, buffer, frames);
if (rc == -EPIPE) {
/* EPIPE means overrun */
fprintf(stderr, "overrun occurred\n");
snd_pcm_prepare(handle);
} else if (rc < 0) {
fprintf(stderr,
"error from read: %s\n",
snd_strerror(rc));
} else if (rc != (int)frames) {
fprintf(stderr, "short read, read %d frames\n", rc);
}
rc = write(1, buffer, size);
if (rc != size)
fprintf(stderr,
"short write: wrote %d bytes\n", rc);
}
snd_pcm_drain(handle);
snd_pcm_close(handle);
free(buffer);
return 0;
}
So, I record a raw file. Then do speech-to-test on that file. But the accuracy is very vert poor. Just like hello or go home will give me hotel or MHM MHM and so on. So what's wrong with these code?
I have read the faqs, should I use acoustic model adaptation to improve accuracy?
PS. I change stereo to mono. And the sound is strange. I can't understand what I said. So, what's wrong with it? This is that raw file test.raw

Answer 1

If you look at the first Q and A in http://cmusphinx.sourceforge.net/wiki/faq you will notice that the library assumes mono data.
You record in stereo.