void
acmod_start_stream(acmod_t *acmod)
{
fe_start_stream(acmod->fe);
acmod->utt_start_frame = 0;
}
void
segment_audio()
{
FILE *file;
int16 pcm_buf[BLOCKSIZE];
mfcc_t **cep_buf;
int16 voiced_buf = NULL;
int32 voiced_nsamps, out_frameidx, uttstart = 0;
char file_name[1024];
uint8 cur_vad_state, vad_state, writing;
int uttno, uttlen, sample_rate;
int32 nframes, nframes_tmp;
int16 frame_size, frame_shift, frame_rate;
size_t k;
sample_rate = (int) cmd_ln_float32_r(config, "-samprate");
frame_rate = cmd_ln_int32_r(config, "-frate");
frame_size =
(int32) (cmd_ln_float32_r(config, "-wlen") * sample_rate + 0.5);
frame_shift =
(int32) (sample_rate / cmd_ln_int32_r(config, "-frate") + 0.5);
nframes = (BLOCKSIZE - frame_size) / frame_shift;
cep_buf =
(mfcc_t **) ckd_calloc_2d(nframes, fe_get_output_size(fe),
sizeof(mfcc_t));
uttno = 0;
uttlen = 0;
cur_vad_state = 0;
voiced_nsamps = 0;
writing = 0;
file = NULL;
fe_start_stream(fe);
fe_start_utt(fe);
while ((k = read_audio(pcm_buf, BLOCKSIZE)) > 0) {
int16 const *pcm_buf_tmp;
pcm_buf_tmp = &pcm_buf[0];
while (k) {
nframes_tmp = nframes;
fe_process_frames_ext(fe, &pcm_buf_tmp, &k, cep_buf,
&nframes_tmp, voiced_buf,
&voiced_nsamps, &out_frameidx);
if (out_frameidx > 0) {
uttstart = out_frameidx;
}
vad_state = fe_get_vad_state(fe);
if (!cur_vad_state && vad_state) {
/* silence->speech transition, time to start new file */
uttno++;
if (!singlefile) {
sprintf(file_name, "%s%04d.raw", infile_path, uttno);
if ((file = fopen(file_name, "wb")) == NULL)
E_FATAL_SYSTEM("Failed to open '%s' for writing",
file_name);
} else {
sprintf(file_name, "%s.raw", infile_path);
if ((file = fopen(file_name, "ab")) == NULL)
E_FATAL_SYSTEM("Failed to open '%s' for writing",
file_name);
}
writing = 1;
}
if (writing && file && voiced_nsamps > 0) {
fwrite(voiced_buf, sizeof(int16), voiced_nsamps, file);
uttlen += voiced_nsamps;
}
if (cur_vad_state && !vad_state) {
/* speech -> silence transition, time to finish file */
fclose(file);
printf("Utterance %04d: file %s start %.1f sec length %d samples ( %.2f sec )\n",
uttno,
file_name,
((double) uttstart) / frame_rate,
uttlen,
((double) uttlen) / sample_rate);
fflush(stdout);
fe_end_utt(fe, cep_buf[0], &nframes_tmp);
writing = 0;
uttlen = 0;
voiced_nsamps = 0;
fe_start_utt(fe);
}
cur_vad_state = vad_state;
}
}
if (writing) {
fclose(file);
printf("Utterance %04d: file %s start %.1f sec length %d samples ( %.2f sec )\n",
uttno,
file_name,
((double) uttstart) / frame_rate,
uttlen,
((double) uttlen) / sample_rate);
fflush(stdout);
}
fe_end_utt(fe, cep_buf[0], &nframes);
ckd_free_2d(cep_buf);
}
fe_t *
fe_init_auto_r(cmd_ln_t *config)
{
fe_t *fe;
int prespch_frame_len;
fe = (fe_t*)ckd_calloc(1, sizeof(*fe));
fe->refcount = 1;
/* transfer params to front end */
if (fe_parse_general_params(cmd_ln_retain(config), fe) < 0) {
fe_free(fe);
return NULL;
}
/* compute remaining fe parameters */
/* We add 0.5 so approximate the float with the closest
* integer. E.g., 2.3 is truncate to 2, whereas 3.7 becomes 4
*/
fe->frame_shift = (int32) (fe->sampling_rate / fe->frame_rate + 0.5);
fe->frame_size = (int32) (fe->window_length * fe->sampling_rate + 0.5);
fe->prior = 0;
fe_start_stream(fe);
assert (fe->frame_shift > 1);
if (fe->frame_size < fe->frame_shift) {
E_ERROR
("Frame size %d (-wlen) must be greater than frame shift %d (-frate)\n",
fe->frame_size, fe->frame_shift);
fe_free(fe);
return NULL;
}
if (fe->frame_size > (fe->fft_size)) {
E_ERROR
("Number of FFT points has to be a power of 2 higher than %d, it is %d\n",
fe->frame_size, fe->fft_size);
fe_free(fe);
return NULL;
}
if (fe->dither)
fe_init_dither(fe->seed);
/* establish buffers for overflow samps and hamming window */
fe->overflow_samps = ckd_calloc(fe->frame_size, sizeof(int16));
fe->hamming_window = ckd_calloc(fe->frame_size/2, sizeof(window_t));
/* create hamming window */
fe_create_hamming(fe->hamming_window, fe->frame_size);
/* init and fill appropriate filter structure */
fe->mel_fb = ckd_calloc(1, sizeof(*fe->mel_fb));
/* transfer params to mel fb */
fe_parse_melfb_params(config, fe, fe->mel_fb);
if (fe->mel_fb->upper_filt_freq > fe->sampling_rate / 2 + 1.0) {
E_ERROR("Upper frequency %.1f is higher than samprate/2 (%.1f)\n",
fe->mel_fb->upper_filt_freq, fe->sampling_rate / 2);
fe_free(fe);
return NULL;
}
fe_build_melfilters(fe->mel_fb);
fe_compute_melcosine(fe->mel_fb);
if (fe->remove_noise || fe->remove_silence)
fe->noise_stats = fe_init_noisestats(fe->mel_fb->num_filters);
fe->vad_data = (vad_data_t*)ckd_calloc(1, sizeof(*fe->vad_data));
prespch_frame_len = fe->log_spec != RAW_LOG_SPEC ? fe->num_cepstra : fe->mel_fb->num_filters;
fe->vad_data->prespch_buf = fe_prespch_init(fe->pre_speech + 1, prespch_frame_len, fe->frame_shift);
/* Create temporary FFT, spectrum and mel-spectrum buffers. */
/* FIXME: Gosh there are a lot of these. */
fe->spch = ckd_calloc(fe->frame_size, sizeof(*fe->spch));
fe->frame = ckd_calloc(fe->fft_size, sizeof(*fe->frame));
fe->spec = ckd_calloc(fe->fft_size, sizeof(*fe->spec));
fe->mfspec = ckd_calloc(fe->mel_fb->num_filters, sizeof(*fe->mfspec));
/* create twiddle factors */
fe->ccc = ckd_calloc(fe->fft_size / 4, sizeof(*fe->ccc));
fe->sss = ckd_calloc(fe->fft_size / 4, sizeof(*fe->sss));
fe_create_twiddle(fe);
if (cmd_ln_boolean_r(config, "-verbose")) {
fe_print_current(fe);
}
/*** Initialize the overflow buffers ***/
fe_start_utt(fe);
return fe;
}
