sonusai 0.17.0__py3-none-any.whl → 0.17.2__py3-none-any.whl

sonusai/calc_metric_spenh.py

@@ -1,27 +1,25 @@
 """sonusai calc_metric_spenh
 
-usage: calc_metric_spenh [-hvtpws] [-i MIXID] [-e WER] [-m WMNAME] PLOC TLOC
+usage: calc_metric_spenh [-hvtpws] [-i MIXID] [-e ASR] [-m MODEL] PLOC TLOC
 
 options:
     -h, --help
     -v, --verbose               Be verbose.
-    -i MIXID, --mixid MIXID     Mixture ID(s) to process, can be range like 0:maxmix+1 [default: *].
+    -i MIXID, --mixid MIXID     Mixture ID(s) to process, can be range like 0:maxmix+1. [default: *]
     -t, --truth-est-mode        Calculate extraction and metrics using truth (instead of prediction).
     -p, --plot                  Enable PDF plots file generation per mixture.
     -w, --wav                   Generate WAV files per mixture.
     -s, --summary               Enable summary files generation.
-    -e WER, --wer-method WER    Word-Error-Rate method: deepgram, google, aixplain_whisper
-                                or whisper (locally run) [default: none]
-    -m WMNAME, --whisper-model  Whisper model name used in aixplain_whisper and whisper WER methods.
-                                [default: tiny]
+    -e ASR, --asr-method ASR    ASR method: deepgram, google, aixplain_whisper, whisper, or sensory. [default: none]
+    -m MODEL, --model           ASR model name used in some ASR methods. [default: tiny]
 
-Calculate speech enhancement metrics of prediction data in PLOC using SonusAI mixture data
-in TLOC as truth/label reference. Metric and extraction data files are written into PLOC.
+Calculate speech enhancement metrics of prediction data in PLOC using SonusAI mixture data in TLOC as truth/label
+reference. Metric and extraction data files are written into PLOC.
 
 PLOC  directory containing prediction data in .h5 files created from truth/label mixture data in TLOC
 TLOC  directory with SonusAI mixture database of truth/label mixture data
 
-For whisper WER methods, the possible models used in local processing (WER = whisper) are:
+For whisper ASR methods, the possible models used in local processing (ASR = whisper) are:
     {tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large}
 but note most are very computationally demanding and can overwhelm/hang a local system.
 
@@ -68,6 +66,7 @@ import matplotlib
 import matplotlib.pyplot as plt
 import numpy as np
 import pandas as pd
+
 from sonusai.mixture import AudioF
 from sonusai.mixture import AudioT
 from sonusai.mixture import Feature
@@ -93,12 +92,12 @@ matplotlib.use('SVG')
 class MPGlobal:
     mixdb: MixtureDatabase = None
     predict_location: str = None
-    predwav_mode: bool = None
+    predict_wav_mode: bool = None
     truth_est_mode: bool = None
     enable_plot: bool = None
     enable_wav: bool = None
-    wer_method: str = None
-    whisper_model: str = None
+    asr_method: str = None
+    asr_model_name: str = None
 
 
 MP_GLOBAL = MPGlobal()
@@ -132,64 +131,62 @@ def snr(clean_speech, processed_speech, sample_rate):
     overall_snr = 10 * np.log10(np.sum(np.square(clean_speech)) / np.sum(np.square(clean_speech - processed_speech)))
 
     # Global Variables
-    winlength = round(30 * sample_rate / 1000)  # window length in samples
-    skiprate = int(np.floor(winlength / 4))  # window skip in samples
-    MIN_SNR = -10  # minimum SNR in dB
-    MAX_SNR = 35  # maximum SNR in dB
+    win_length = round(30 * sample_rate / 1000)  # window length in samples
+    skip_rate = int(np.floor(win_length / 4))  # window skip in samples
+    min_snr = -10  # minimum SNR in dB
+    max_snr = 35  # maximum SNR in dB
 
     # For each frame of input speech, calculate the Segmental SNR
-    num_frames = int(clean_length / skiprate - (winlength / skiprate))  # number of frames
+    num_frames = int(clean_length / skip_rate - (win_length / skip_rate))  # number of frames
     start = 0  # starting sample
-    window = 0.5 * (1 - np.cos(2 * np.pi * np.arange(1, winlength + 1) / (winlength + 1)))
+    window = 0.5 * (1 - np.cos(2 * np.pi * np.arange(1, win_length + 1) / (win_length + 1)))
 
     segmental_snr = np.empty(num_frames)
-    EPS = np.spacing(1)
+    eps = np.spacing(1)
     for frame_count in range(num_frames):
         # (1) Get the Frames for the test and reference speech. Multiply by Hanning Window.
-        clean_frame = clean_speech[start:start + winlength]
-        processed_frame = processed_speech[start:start + winlength]
+        clean_frame = clean_speech[start:start + win_length]
+        processed_frame = processed_speech[start:start + win_length]
         clean_frame = np.multiply(clean_frame, window)
         processed_frame = np.multiply(processed_frame, window)
 
         # (2) Compute the Segmental SNR
         signal_energy = np.sum(np.square(clean_frame))
         noise_energy = np.sum(np.square(clean_frame - processed_frame))
-        segmental_snr[frame_count] = 10 * np.log10(signal_energy / (noise_energy + EPS) + EPS)
-        segmental_snr[frame_count] = max(segmental_snr[frame_count], MIN_SNR)
-        segmental_snr[frame_count] = min(segmental_snr[frame_count], MAX_SNR)
+        segmental_snr[frame_count] = 10 * np.log10(signal_energy / (noise_energy + eps) + eps)
+        segmental_snr[frame_count] = np.max(segmental_snr[frame_count], min_snr)
+        segmental_snr[frame_count] = np.min(segmental_snr[frame_count], max_snr)
 
-        start = start + skiprate
+        start = start + skip_rate
 
     return overall_snr, segmental_snr
 
 
-def lpcoeff(speech_frame, model_order):
+def lp_coefficients(speech_frame, model_order):
     # (1) Compute Autocorrelation Lags
-    winlength = np.size(speech_frame)
-    R = np.empty(model_order + 1)
-    E = np.empty(model_order + 1)
+    win_length = np.size(speech_frame)
+    autocorrelation = np.empty(model_order + 1)
+    e = np.empty(model_order + 1)
     for k in range(model_order + 1):
-        R[k] = np.dot(speech_frame[0:winlength - k], speech_frame[k: winlength])
+        autocorrelation[k] = np.dot(speech_frame[0:win_length - k], speech_frame[k: win_length])
 
     # (2) Levinson-Durbin
     a = np.ones(model_order)
     a_past = np.empty(model_order)
-    rcoeff = np.empty(model_order)
-    E[0] = R[0]
+    ref_coefficients = np.empty(model_order)
+    e[0] = autocorrelation[0]
     for i in range(model_order):
         a_past[0: i] = a[0: i]
-        sum_term = np.dot(a_past[0: i], R[i:0:-1])
-        rcoeff[i] = (R[i + 1] - sum_term) / E[i]
-        a[i] = rcoeff[i]
+        sum_term = np.dot(a_past[0: i], autocorrelation[i:0:-1])
+        ref_coefficients[i] = (autocorrelation[i + 1] - sum_term) / e[i]
+        a[i] = ref_coefficients[i]
         if i == 0:
-            a[0: i] = a_past[0: i] - np.multiply(a_past[i - 1:-1:-1], rcoeff[i])
+            a[0: i] = a_past[0: i] - np.multiply(a_past[i - 1:-1:-1], ref_coefficients[i])
         else:
-            a[0: i] = a_past[0: i] - np.multiply(a_past[i - 1::-1], rcoeff[i])
-        E[i + 1] = (1 - rcoeff[i] * rcoeff[i]) * E[i]
-    acorr = R
-    refcoeff = rcoeff
-    lpparams = np.concatenate((np.array([1]), -a))
-    return acorr, refcoeff, lpparams
+            a[0: i] = a_past[0: i] - np.multiply(a_past[i - 1::-1], ref_coefficients[i])
+        e[i + 1] = (1 - ref_coefficients[i] * ref_coefficients[i]) * e[i]
+    lp_params = np.concatenate((np.array([1]), -a))
+    return autocorrelation, ref_coefficients, lp_params
 
 
 def llr(clean_speech, processed_speech, sample_rate):
@@ -199,38 +196,38 @@ def llr(clean_speech, processed_speech, sample_rate):
     clean_length = np.size(clean_speech)
     processed_length = np.size(processed_speech)
     if clean_length != processed_length:
-        raise ValueError('Both Speech Files must be same length.')
+        raise ValueError('Both speech files must be same length.')
 
     # Global Variables
-    winlength = (np.round(30 * sample_rate / 1000)).astype(int)  # window length in samples
-    skiprate = (np.floor(winlength / 4)).astype(int)  # window skip in samples
+    win_length = (np.round(30 * sample_rate / 1000)).astype(int)  # window length in samples
+    skip_rate = (np.floor(win_length / 4)).astype(int)  # window skip in samples
     if sample_rate < 10000:
-        P = 10  # LPC Analysis Order
+        p = 10  # LPC Analysis Order
     else:
-        P = 16  # this could vary depending on sampling frequency.
+        p = 16  # this could vary depending on sampling frequency.
 
     # For each frame of input speech, calculate the Log Likelihood Ratio
-    num_frames = int((clean_length - winlength) / skiprate)  # number of frames
+    num_frames = int((clean_length - win_length) / skip_rate)  # number of frames
     start = 0  # starting sample
-    window = 0.5 * (1 - np.cos(2 * np.pi * np.arange(1, winlength + 1) / (winlength + 1)))
+    window = 0.5 * (1 - np.cos(2 * np.pi * np.arange(1, win_length + 1) / (win_length + 1)))
 
     distortion = np.empty(num_frames)
     for frame_count in range(num_frames):
         # (1) Get the Frames for the test and reference speech. Multiply by Hanning Window.
-        clean_frame = clean_speech[start: start + winlength]
-        processed_frame = processed_speech[start: start + winlength]
+        clean_frame = clean_speech[start: start + win_length]
+        processed_frame = processed_speech[start: start + win_length]
         clean_frame = np.multiply(clean_frame, window)
         processed_frame = np.multiply(processed_frame, window)
 
         # (2) Get the autocorrelation lags and LPC parameters used to compute the LLR measure.
-        R_clean, Ref_clean, A_clean = lpcoeff(clean_frame, P)
-        R_processed, Ref_processed, A_processed = lpcoeff(processed_frame, P)
+        r_clean, ref_clean, a_clean = lp_coefficients(clean_frame, p)
+        r_processed, ref_processed, a_processed = lp_coefficients(processed_frame, p)
 
         # (3) Compute the LLR measure
-        numerator = np.dot(np.matmul(A_processed, toeplitz(R_clean)), A_processed)
-        denominator = np.dot(np.matmul(A_clean, toeplitz(R_clean)), A_clean)
+        numerator = np.dot(np.matmul(a_processed, toeplitz(r_clean)), a_processed)
+        denominator = np.dot(np.matmul(a_clean, toeplitz(r_clean)), a_clean)
         distortion[frame_count] = np.log(numerator / denominator)
-        start = start + skiprate
+        start = start + skip_rate
     return distortion
 
 
@@ -244,16 +241,15 @@ def wss(clean_speech, processed_speech, sample_rate):
         raise ValueError('Files must have same length.')
 
     # Global variables
-    winlength = (np.round(30 * sample_rate / 1000)).astype(int)  # window length in samples
-    skiprate = (np.floor(np.divide(winlength, 4))).astype(int)  # window skip in samples
+    win_length = (np.round(30 * sample_rate / 1000)).astype(int)  # window length in samples
+    skip_rate = (np.floor(np.divide(win_length, 4))).astype(int)  # window skip in samples
     max_freq = (np.divide(sample_rate, 2)).astype(int)  # maximum bandwidth
     num_crit = 25  # number of critical bands
 
-    USE_FFT_SPECTRUM = 1  # defaults to 10th order LP spectrum
-    n_fft = (np.power(2, np.ceil(np.log2(2 * winlength)))).astype(int)
-    n_fftby2 = (np.multiply(0.5, n_fft)).astype(int)  # FFT size/2
-    Kmax = 20.0  # value suggested by Klatt, pg 1280
-    Klocmax = 1.0  # value suggested by Klatt, pg 1280
+    n_fft = (np.power(2, np.ceil(np.log2(2 * win_length)))).astype(int)
+    n_fft_by_2 = (np.multiply(0.5, n_fft)).astype(int)  # FFT size/2
+    k_max = 20.0  # value suggested by Klatt, pg 1280
+    k_loc_max = 1.0  # value suggested by Klatt, pg 1280
 
     # Critical Band Filter Definitions (Center Frequency and Bandwidths in Hz)
     cent_freq = np.array([50.0000, 120.000, 190.000, 260.000, 330.000, 400.000, 470.000,
@@ -268,39 +264,38 @@ def wss(clean_speech, processed_speech, sample_rate):
     bw_min = bandwidth[0]  # minimum critical bandwidth
 
     # Set up the critical band filters.
-    # Note here that Gaussianly shaped filters are used.
+    # Note here that Gaussian-ly shaped filters are used.
     # Also, the sum of the filter weights are equivalent for each critical band filter.
     # Filter less than -30 dB and set to zero.
     min_factor = np.exp(-30.0 / (2.0 * 2.303))  # -30 dB point of filter
-    crit_filter = np.empty((num_crit, n_fftby2))
+    crit_filter = np.empty((num_crit, n_fft_by_2))
     for i in range(num_crit):
-        f0 = (cent_freq[i] / max_freq) * n_fftby2
-        bw = (bandwidth[i] / max_freq) * n_fftby2
+        f0 = (cent_freq[i] / max_freq) * n_fft_by_2
+        bw = (bandwidth[i] / max_freq) * n_fft_by_2
         norm_factor = np.log(bw_min) - np.log(bandwidth[i])
-        j = np.arange(n_fftby2)
+        j = np.arange(n_fft_by_2)
         crit_filter[i, :] = np.exp(-11 * np.square(np.divide(j - np.floor(f0), bw)) + norm_factor)
         cond = np.greater(crit_filter[i, :], min_factor)
         crit_filter[i, :] = np.where(cond, crit_filter[i, :], 0)
     # For each frame of input speech, calculate the Weighted Spectral Slope Measure
-    num_frames = int(clean_length / skiprate - (winlength / skiprate))  # number of frames
+    num_frames = int(clean_length / skip_rate - (win_length / skip_rate))  # number of frames
     start = 0  # starting sample
-    window = 0.5 * (1 - np.cos(2 * np.pi * np.arange(1, winlength + 1) / (winlength + 1)))
+    window = 0.5 * (1 - np.cos(2 * np.pi * np.arange(1, win_length + 1) / (win_length + 1)))
 
     distortion = np.empty(num_frames)
     for frame_count in range(num_frames):
         # (1) Get the Frames for the test and reference speech. Multiply by Hanning Window.
-        clean_frame = clean_speech[start: start + winlength] / 32768
-        processed_frame = processed_speech[start: start + winlength] / 32768
+        clean_frame = clean_speech[start: start + win_length] / 32768
+        processed_frame = processed_speech[start: start + win_length] / 32768
         clean_frame = np.multiply(clean_frame, window)
         processed_frame = np.multiply(processed_frame, window)
         # (2) Compute the Power Spectrum of Clean and Processed
-        # if USE_FFT_SPECTRUM:
         clean_spec = np.square(np.abs(fft(clean_frame, n_fft)))
         processed_spec = np.square(np.abs(fft(processed_frame, n_fft)))
 
         # (3) Compute Filterbank Output Energies (in dB scale)
-        clean_energy = np.matmul(crit_filter, clean_spec[0:n_fftby2])
-        processed_energy = np.matmul(crit_filter, processed_spec[0:n_fftby2])
+        clean_energy = np.matmul(crit_filter, clean_spec[0:n_fft_by_2])
+        processed_energy = np.matmul(crit_filter, processed_spec[0:n_fft_by_2])
 
         clean_energy = 10 * np.log10(np.maximum(clean_energy, 1E-10))
         processed_energy = 10 * np.log10(np.maximum(processed_energy, 1E-10))
@@ -340,39 +335,39 @@ def wss(clean_speech, processed_speech, sample_rate):
                 processed_loc_peak[i] = processed_energy[n + 1]
 
         # (6) Compute the WSS Measure for this frame. This includes determination of the weighting function.
-        dBMax_clean = np.max(clean_energy)
-        dBMax_processed = np.max(processed_energy)
+        db_max_clean = np.max(clean_energy)
+        db_max_processed = np.max(processed_energy)
         '''
         The weights are calculated by averaging individual weighting factors from the clean and processed frame.
-        These weights W_clean and W_processed should range from 0 to 1 and place more emphasis on spectral peaks
+        These weights w_clean and w_processed should range from 0 to 1 and place more emphasis on spectral peaks
         and less emphasis on slope differences in spectral valleys.
         This procedure is described on page 1280 of Klatt's 1982 ICASSP paper.
         '''
-        Wmax_clean = np.divide(Kmax, Kmax + dBMax_clean - clean_energy[0: num_crit - 1])
-        Wlocmax_clean = np.divide(Klocmax, Klocmax + clean_loc_peak - clean_energy[0: num_crit - 1])
-        W_clean = np.multiply(Wmax_clean, Wlocmax_clean)
+        w_max_clean = np.divide(k_max, k_max + db_max_clean - clean_energy[0: num_crit - 1])
+        w_loc_max_clean = np.divide(k_loc_max, k_loc_max + clean_loc_peak - clean_energy[0: num_crit - 1])
+        w_clean = np.multiply(w_max_clean, w_loc_max_clean)
 
-        Wmax_processed = np.divide(Kmax, Kmax + dBMax_processed - processed_energy[0: num_crit - 1])
-        Wlocmax_processed = np.divide(Klocmax, Klocmax + processed_loc_peak - processed_energy[0: num_crit - 1])
-        W_processed = np.multiply(Wmax_processed, Wlocmax_processed)
+        w_max_processed = np.divide(k_max, k_max + db_max_processed - processed_energy[0: num_crit - 1])
+        w_loc_max_processed = np.divide(k_loc_max, k_loc_max + processed_loc_peak - processed_energy[0: num_crit - 1])
+        w_processed = np.multiply(w_max_processed, w_loc_max_processed)
 
-        W = np.divide(np.add(W_clean, W_processed), 2.0)
+        w = np.divide(np.add(w_clean, w_processed), 2.0)
         slope_diff = np.subtract(clean_slope, processed_slope)[0: num_crit - 1]
-        distortion[frame_count] = np.dot(W, np.square(slope_diff)) / np.sum(W)
-        # this normalization is not part of Klatt's paper, but helps to normalize the measure.
+        distortion[frame_count] = np.dot(w, np.square(slope_diff)) / np.sum(w)
+        # This normalization is not part of Klatt's paper, but helps to normalize the measure.
         # Here we scale the measure by the sum of the weights.
-        start = start + skiprate
+        start = start + skip_rate
     return distortion
 
 
 def calc_speech_metrics(hypothesis: np.ndarray,
                         reference: np.ndarray) -> tuple[float, int, int, int, float]:
     """
-    Calculate speech metrics pesq_mos, CSIG, CBAK, COVL, segSNR.  These are all related and thus included
+    Calculate speech metrics pesq_mos, c_sig, c_bak, c_ovl, seg_snr.  These are all related and thus included
     in one function. Reference: matlab script "compute_metrics.m".
 
     Usage:
-        pesq, csig, cbak, covl, ssnr = compute_metrics(hypothesis, reference, Fs, path)
+        pesq, c_sig, c_bak, c_ovl, ssnr = compute_metrics(hypothesis, reference, fs, path)
         reference: clean audio as array
         hypothesis: enhanced audio as array
         Audio must have sampling rate = 16000 Hz.
@@ -383,41 +378,39 @@ def calc_speech_metrics(hypothesis: np.ndarray,
     """
     from sonusai.metrics import calc_pesq
 
-    Fs = 16000
+    fs = 16000
 
     # compute the WSS measure
-    wss_dist_vec = wss(reference, hypothesis, Fs)
+    wss_dist_vec = wss(reference, hypothesis, fs)
     wss_dist_vec = np.sort(wss_dist_vec)
     alpha = 0.95  # value from CMGAN ref implementation
     wss_dist = np.mean(wss_dist_vec[0: round(np.size(wss_dist_vec) * alpha)])
 
     # compute the LLR measure
-    llr_dist = llr(reference, hypothesis, Fs)
+    llr_dist = llr(reference, hypothesis, fs)
     ll_rs = np.sort(llr_dist)
     llr_len = round(np.size(llr_dist) * alpha)
     llr_mean = np.mean(ll_rs[0: llr_len])
 
     # compute the SNRseg
-    snr_dist, segsnr_dist = snr(reference, hypothesis, Fs)
-    snr_mean = snr_dist
-    segSNR = np.mean(segsnr_dist)
+    snr_dist, segsnr_dist = snr(reference, hypothesis, fs)
+    seg_snr = np.mean(segsnr_dist)
 
     # compute the pesq (use Sonusai wrapper, only fs=16k, mode=wb support)
     pesq_mos = calc_pesq(hypothesis=hypothesis, reference=reference)
-    # pesq_mos = pesq(sampling_rate1, data1, data2, 'wb')
 
     # now compute the composite measures
-    CSIG = 3.093 - 1.029 * llr_mean + 0.603 * pesq_mos - 0.009 * wss_dist
-    CSIG = max(1, CSIG)
-    CSIG = min(5, CSIG)  # limit values to [1, 5]
-    CBAK = 1.634 + 0.478 * pesq_mos - 0.007 * wss_dist + 0.063 * segSNR
-    CBAK = max(1, CBAK)
-    CBAK = min(5, CBAK)  # limit values to [1, 5]
-    COVL = 1.594 + 0.805 * pesq_mos - 0.512 * llr_mean - 0.007 * wss_dist
-    COVL = max(1, COVL)
-    COVL = min(5, COVL)  # limit values to [1, 5]
+    c_sig = 3.093 - 1.029 * llr_mean + 0.603 * pesq_mos - 0.009 * wss_dist
+    c_sig = max(1, c_sig)
+    c_sig = min(5, c_sig)  # limit values to [1, 5]
+    c_bak = 1.634 + 0.478 * pesq_mos - 0.007 * wss_dist + 0.063 * seg_snr
+    c_bak = max(1, c_bak)
+    c_bak = min(5, c_bak)  # limit values to [1, 5]
+    c_ovl = 1.594 + 0.805 * pesq_mos - 0.512 * llr_mean - 0.007 * wss_dist
+    c_ovl = max(1, c_ovl)
+    c_ovl = min(5, c_ovl)  # limit values to [1, 5]
 
-    return pesq_mos, CSIG, CBAK, COVL, segSNR
+    return pesq_mos, c_sig, c_bak, c_ovl, seg_snr
 
 
 def mean_square_error(hypothesis: np.ndarray,
@@ -564,10 +557,8 @@ def plot_mixpred(mixture: AudioT,
     ax[p].plot(x_axis, mixture, label='Mixture', color='mistyrose')
     ax[0].set_ylabel('magnitude', color='tab:blue')
     ax[p].set_xlim(x_axis[0], x_axis[-1])
-    # ax[p].set_ylim([-1.025, 1.025])
     if target is not None:  # Plot target time-domain waveform on top of mixture
         ax[0].plot(x_axis, target, label='Target', color='tab:blue')
-        # ax[0].tick_params(axis='y', labelcolor=color)
     ax[p].set_title('Waveform')
 
     # Plot the mixture spectrogram
@@ -589,10 +580,10 @@ def plot_mixpred(mixture: AudioT,
     return fig
 
 
-def plot_pdb_predtruth(predict: np.ndarray,
-                       truth_f: Optional[np.ndarray] = None,
-                       metric: Optional[np.ndarray] = None,
-                       tp_title: str = '') -> plt.Figure:
+def plot_pdb_predict_truth(predict: np.ndarray,
+                           truth_f: Optional[np.ndarray] = None,
+                           metric: Optional[np.ndarray] = None,
+                           tp_title: str = '') -> plt.Figure:
     """Plot predict and optionally truth and a metric in power db, e.g. applies 10*log10(predict)"""
     num_plots = 2
     if truth_f is not None:
@@ -636,16 +627,15 @@ def plot_pdb_predtruth(predict: np.ndarray,
         ax[p].set_title('SNR and SNR mse (mean over freq. db)')
     else:
         ax[p].set_title('SNR (mean over freq. db)')
-        # ax[0].tick_params(axis='y', labelcolor=color)
     return fig
 
 
-def plot_epredtruth(predict: np.ndarray,
-                    predict_wav: np.ndarray,
-                    truth_f: Optional[np.ndarray] = None,
-                    truth_wav: Optional[np.ndarray] = None,
-                    metric: Optional[np.ndarray] = None,
-                    tp_title: str = '') -> plt.Figure:
+def plot_e_predict_truth(predict: np.ndarray,
+                         predict_wav: np.ndarray,
+                         truth_f: Optional[np.ndarray] = None,
+                         truth_wav: Optional[np.ndarray] = None,
+                         metric: Optional[np.ndarray] = None,
+                         tp_title: str = '') -> plt.Figure:
     """Plot predict spectrogram and waveform and optionally truth and a metric)"""
     num_plots = 2
     if truth_f is not None:
@@ -666,7 +656,7 @@ def plot_epredtruth(predict: np.ndarray,
         ax[p].imshow(truth_f.transpose(), im.cmap, aspect='auto', interpolation='nearest', origin='lower')
         ax[p].set_title('Truth')
 
-    # Plot the predict wav, and optionally truth avg and metric lines
+    # Plot predict wav, and optionally truth avg and metric lines
     p += 1
     x_axis = np.arange(len(predict_wav), dtype=np.float32)  # / SAMPLE_RATE
     ax[p].plot(x_axis, predict_wav, color='black', linestyle='dashed', label='Speech Estimate')
@@ -732,12 +722,12 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
 
     mixdb = MP_GLOBAL.mixdb
     predict_location = MP_GLOBAL.predict_location
-    predwav_mode = MP_GLOBAL.predwav_mode
+    predict_wav_mode = MP_GLOBAL.predict_wav_mode
     truth_est_mode = MP_GLOBAL.truth_est_mode
     enable_plot = MP_GLOBAL.enable_plot
     enable_wav = MP_GLOBAL.enable_wav
-    wer_method = MP_GLOBAL.wer_method
-    whisper_model = MP_GLOBAL.whisper_model
+    asr_method = MP_GLOBAL.asr_method
+    asr_model_name = MP_GLOBAL.asr_model_name
 
     # 1)  Read predict data, var predict with shape [BatchSize,Classes] or [BatchSize,Tsteps,Classes]
     output_name = join(predict_location, mixdb.mixture(mixid).name)
@@ -749,7 +739,7 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
         base_name = splitext(output_name)[0] + '_truest'
     else:
         base_name, ext = splitext(output_name)  # base_name used later
-        if not predwav_mode:
+        if not predict_wav_mode:
             try:
                 with h5py.File(output_name, 'r') as f:
                     predict = np.array(f['predict'])
@@ -761,8 +751,8 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
                 predict, _ = reshape_outputs(predict=predict, truth=None, timesteps=predict.shape[1])
         else:
             base_name, ext = splitext(output_name)
-            prfname = join(base_name + '.wav')
-            audio = read_audio(prfname)
+            predict_name = join(base_name + '.wav')
+            audio = read_audio(predict_name)
             predict = forward_transform(audio, mixdb.ft_config)
             if mixdb.feature[0:1] == 'h':
                 predict = power_compress(predict)
@@ -773,8 +763,8 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
     target_f = mixdb.mixture_targets_f(mixid, targets=tmp)[0]
     target = tmp[0]
     mixture = mixdb.mixture_mixture(mixid)  # note: gives full reverberated/distorted target, but no specaugment
-    # noise_wodist = mixdb.mixture_noise(mixid)            # noise without specaugment and distortion
-    # noise_wodist_f = mixdb.mixture_noise_f(mixid, noise=noise_wodist)
+    # noise_wo_dist = mixdb.mixture_noise(mixid)            # noise without specaugment and distortion
+    # noise_wo_dist_f = mixdb.mixture_noise_f(mixid, noise=noise_wo_dist)
     noise = mixture - target  # has time-domain distortion (ir,etc.) but does not have specaugment
     # noise_f = mixdb.mixture_noise_f(mixid, noise=noise)
     segsnr_f = mixdb.mixture_segsnr(mixid, target=target, noise=noise)  # note: uses pre-IR, pre-specaug audio
@@ -784,9 +774,9 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
     segsnr_f[segsnr_f == inf] = 7.944e8  # 99db
     segsnr_f[segsnr_f == -inf] = 1.258e-10  # -99db
     # need to use inv-tf to match #samples & latency shift properties of predict inv tf
-    targetfi = inverse_transform(target_f, mixdb.it_config)
-    noisefi = inverse_transform(noise_f, mixdb.it_config)
-    # mixturefi = mixdb.inverse_transform(mixture_f)
+    target_fi = inverse_transform(target_f, mixdb.it_config)
+    noise_fi = inverse_transform(noise_f, mixdb.it_config)
+    # mixture_fi = mixdb.inverse_transform(mixture_f)
 
     # gen feature, truth - note feature only used for plots
     # TBD parse truth_f for different formats and also multi-truth
@@ -798,17 +788,17 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
 
     if not truth_est_mode:
         if predict.shape[0] < target_f.shape[0]:  # target_f, truth_f, mixture_f, etc. same size
-            trimf = target_f.shape[0] - predict.shape[0]
-            logger.debug(f'Warning: prediction frames less than mixture, trimming {trimf} frames from all truth.')
-            target_f = target_f[0:-trimf, :]
-            targetfi, _ = inverse_transform(target_f, mixdb.it_config)
-            trimt = target.shape[0] - targetfi.shape[0]
-            target = target[0:-trimt]
-            noise_f = noise_f[0:-trimf, :]
-            noise = noise[0:-trimt]
-            mixture_f = mixture_f[0:-trimf, :]
-            mixture = mixture[0:-trimt]
-            truth_f = truth_f[0:-trimf, :]
+            trim_f = target_f.shape[0] - predict.shape[0]
+            logger.debug(f'Warning: prediction frames less than mixture, trimming {trim_f} frames from all truth.')
+            target_f = target_f[0:-trim_f, :]
+            target_fi, _ = inverse_transform(target_f, mixdb.it_config)
+            trim_t = target.shape[0] - target_fi.shape[0]
+            target = target[0:-trim_t]
+            noise_f = noise_f[0:-trim_f, :]
+            noise = noise[0:-trim_t]
+            mixture_f = mixture_f[0:-trim_f, :]
+            mixture = mixture[0:-trim_t]
+            truth_f = truth_f[0:-trim_f, :]
         elif predict.shape[0] > target_f.shape[0]:
             raise SonusAIError(
                 f'Error: prediction has more frames than true mixture {predict.shape[0]} vs {truth_f.shape[0]}')
@@ -848,14 +838,14 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
     phd, phd_bin, phd_frame = phase_distance(hypothesis=predict_complex, reference=truth_f_complex)
 
     # Noise td logerr
-    # lerr_nt, lerr_nt_bin, lerr_nt_frame = log_error(noisefi, noise_truth_est_audio)
+    # lerr_nt, lerr_nt_bin, lerr_nt_frame = log_error(noise_fi, noise_truth_est_audio)
 
     # # SA-SDR (time-domain source-aggragated SDR)
-    ytrue = np.concatenate((targetfi[:, np.newaxis], noisefi[:, np.newaxis]), axis=1)
+    ytrue = np.concatenate((target_fi[:, np.newaxis], noise_fi[:, np.newaxis]), axis=1)
     ypred = np.concatenate((target_est_wav[:, np.newaxis], noise_est_wav[:, np.newaxis]), axis=1)
     # # note: w/o scale is more pessimistic number
     # sa_sdr, _ = calc_sa_sdr(hypothesis=ypred, reference=ytrue)
-    target_stoi = stoi(targetfi, target_est_wav, 16000, extended=False)
+    target_stoi = stoi(target_fi, target_est_wav, 16000, extended=False)
 
     wsdr, wsdr_cc, wsdr_cw = calc_wsdr(hypothesis=ypred, reference=ytrue, with_log=True)
     # logger.debug(f'wsdr weight sum for mixid {mixid} = {np.sum(wsdr_cw)}.')
@@ -865,7 +855,7 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
     # Speech intelligibility measure - PESQ
     if int(mixdb.mixture(mixid).snr) > -99:
         # len = target_est_wav.shape[0]
-        pesq_speech, csig_tg, cbak_tg, covl_tg, sgsnr_tg = calc_speech_metrics(target_est_wav, targetfi)
+        pesq_speech, csig_tg, cbak_tg, covl_tg, sgsnr_tg = calc_speech_metrics(target_est_wav, target_fi)
         pesq_mixture, csig_mx, cbak_mx, covl_mx, sgsnr_mx = calc_speech_metrics(mixture, target)
         # pesq_speech_tst = calc_pesq(hypothesis=target_est_wav, reference=target)
         # pesq_mixture_tst = calc_pesq(hypothesis=mixture, reference=target)
@@ -884,23 +874,26 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
         covl_mx = 0
         covl_tg = 0
 
-    # Calc WER
-    asr_tt = ''
-    asr_mx = ''
-    asr_tge = ''
-    if wer_method == 'none' or mixdb.mixture(mixid).snr == -99:  # noise only, ignore/reset target asr
+    # Calc ASR
+    asr_tt = None
+    asr_mx = None
+    asr_tge = None
+    if asr_method == 'none' or mixdb.mixture(mixid).snr == -99:  # noise only, ignore/reset target asr
         wer_mx = float('nan')
         wer_tge = float('nan')
         wer_pi = float('nan')
     else:
-        if MP_GLOBAL.mixdb.asr_manifests:
-            asr_tt = MP_GLOBAL.mixdb.mixture_asr_data(mixid)[0]  # ignore mixup
-        else:
-            asr_tt = calc_asr(target, engine=wer_method, whisper_model_name=whisper_model).text  # target truth
+        asr_tt = MP_GLOBAL.mixdb.get_speech_metadata(mixid, 'text')[0]  # ignore mixup
+        if asr_tt is None:
+            asr_tt = calc_asr(target, engine=asr_method, whisper_model_name=asr_model_name).text  # target truth
+        # if MP_GLOBAL.mixdb.asr_manifests:
+        #     asr_tt = MP_GLOBAL.mixdb.mixture_asr_data(mixid)[0]  # ignore mixup
+        # else:
+        #     asr_tt = calc_asr(target, engine=asr_method, whisper_model_name=asr_model_name).text  # target truth
 
         if asr_tt:
-            asr_mx = calc_asr(mixture, engine=wer_method, whisper_model=whisper_model).text
-            asr_tge = calc_asr(target_est_wav, engine=wer_method, whisper_model=whisper_model).text
+            asr_mx = calc_asr(mixture, engine=asr_method, whisper_model_name=asr_model_name).text
+            asr_tge = calc_asr(target_est_wav, engine=asr_method, whisper_model_name=asr_model_name).text
 
             wer_mx = calc_wer(asr_mx, asr_tt).wer * 100  # mixture wer
             wer_tge = calc_wer(asr_tge, asr_tt).wer * 100  # target estimate wer
@@ -962,10 +955,10 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
         print('', file=f)
         print(f'Target path: {mixdb.target_file(ti).name}', file=f)
         print(f'Noise path: {mixdb.noise_file(ni).name}', file=f)
-        if wer_method != 'none':
-            print(f'WER method: {wer_method} and whisper model (if used):  {whisper_model}', file=f)
+        if asr_method != 'none':
+            print(f'ASR method: {asr_method} and whisper model (if used):  {asr_model_name}', file=f)
             if mixdb.asr_manifests:
-                print(f'ASR truth from manifest:  {asr_tt}', file=f)
+                print(f'ASR truth from metadata:  {asr_tt}', file=f)
             else:
                 print(f'ASR truth from wer method:  {asr_tt}', file=f)
             print(f'ASR result for mixture:  {asr_mx}', file=f)
@@ -977,7 +970,7 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
     if enable_wav:
         write_wav(name=base_name + '_mixture.wav', audio=float_to_int16(mixture))
         write_wav(name=base_name + '_target.wav', audio=float_to_int16(target))
-        # write_wav(name=base_name + '_targetfi.wav', audio=float_to_int16(targetfi))
+        # write_wav(name=base_name + '_target_fi.wav', audio=float_to_int16(target_fi))
         write_wav(name=base_name + '_noise.wav', audio=float_to_int16(noise))
         write_wav(name=base_name + '_target_est.wav', audio=float_to_int16(target_est_wav))
         write_wav(name=base_name + '_noise_est.wav', audio=float_to_int16(noise_est_wav))
@@ -992,7 +985,7 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
     # 8) Write out plot file
     if enable_plot:
         from matplotlib.backends.backend_pdf import PdfPages
-        plot_fname = base_name + '_metric_spenh.pdf'
+        plot_name = base_name + '_metric_spenh.pdf'
 
         # Reshape feature to eliminate overlap redundancy for easier to understand spectrogram view
         # Original size (frames, stride, num_bands), decimates in stride dimension only if step is > 1
@@ -1007,7 +1000,7 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
         feat_sgram = feat_sgram[:, -step:, :]  # decimate,  Fx1xB
         feat_sgram = np.reshape(feat_sgram, (feat_sgram.shape[0] * feat_sgram.shape[1], feat_sgram.shape[2]))
 
-        with PdfPages(plot_fname) as pdf:
+        with PdfPages(plot_name) as pdf:
             # page1 we always have a mixture and prediction, target optional if truth provided
             tfunc_name = mixdb.target_file(1).truth_settings[0].function  # first target, assumes all have same
             if tfunc_name == 'mapped_snr_f':
@@ -1036,25 +1029,25 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
             tg_spec = 20 * np.log10(abs(target_f) + np.finfo(np.float32).eps)
             tg_est_spec = 20 * np.log10(abs(predict_complex) + np.finfo(np.float32).eps)
             # n_spec = np.reshape(n_spec,(n_spec.shape[0] * n_spec.shape[1], n_spec.shape[2]))
-            pdf.savefig(plot_epredtruth(predict=tg_est_spec,
-                                        predict_wav=target_est_wav,
-                                        truth_f=tg_spec,
-                                        truth_wav=targetfi,
-                                        metric=np.vstack((lerr_tg_frame, phd_frame)).T,
-                                        tp_title='speech estimate'))
+            pdf.savefig(plot_e_predict_truth(predict=tg_est_spec,
+                                             predict_wav=target_est_wav,
+                                             truth_f=tg_spec,
+                                             truth_wav=target_fi,
+                                             metric=np.vstack((lerr_tg_frame, phd_frame)).T,
+                                             tp_title='speech estimate'))
 
             # page 4 noise extraction
             n_spec = 20 * np.log10(abs(noise_f) + np.finfo(np.float32).eps)
             n_est_spec = 20 * np.log10(abs(noise_est_complex) + np.finfo(np.float32).eps)
-            pdf.savefig(plot_epredtruth(predict=n_est_spec,
-                                        predict_wav=noise_est_wav,
-                                        truth_f=n_spec,
-                                        truth_wav=noisefi,
-                                        metric=lerr_n_frame,
-                                        tp_title='noise estimate'))
+            pdf.savefig(plot_e_predict_truth(predict=n_est_spec,
+                                             predict_wav=noise_est_wav,
+                                             truth_f=n_spec,
+                                             truth_wav=noise_fi,
+                                             metric=lerr_n_frame,
+                                             tp_title='noise estimate'))
 
             # Plot error waveforms
-            # tg_err_wav = targetfi - target_est_wav
+            # tg_err_wav = target_fi - target_est_wav
             # tg_err_spec = 20*np.log10(np.abs(target_f - predict_complex))
 
         plt.close('all')
@@ -1072,14 +1065,14 @@ def main():
 
     verbose = args['--verbose']
     mixids = args['--mixid']
-    predict_location = args['PLOC']
-    wer_method = args['--wer-method'].lower()
+    asr_method = args['--asr-method'].lower()
+    asr_model_name = args['--model'].lower()
     truth_est_mode = args['--truth-est-mode']
     enable_plot = args['--plot']
     enable_wav = args['--wav']
     enable_summary = args['--summary']
+    predict_location = args['PLOC']
     truth_location = args['TLOC']
-    whisper_model = args['--whisper-model'].lower()
 
     import glob
     from os.path import basename
@@ -1103,19 +1096,19 @@ def main():
     if not isdir(predict_location):
         print(f'The specified predict location {predict_location} is not a valid subdirectory path, exiting ...')
 
-    # allpfiles = listdir(predict_location)
-    allpfiles = glob.glob(predict_location + "/*.h5")
+    # all_predict_files = listdir(predict_location)
+    all_predict_files = glob.glob(predict_location + "/*.h5")
     predict_logfile = glob.glob(predict_location + "/*predict.log")
-    predwav_mode = False
-    if len(allpfiles) <= 0 and not truth_est_mode:
-        allpfiles = glob.glob(predict_location + "/*.wav")  # check for wav files
-        if len(allpfiles) <= 0:
+    predict_wav_mode = False
+    if len(all_predict_files) <= 0 and not truth_est_mode:
+        all_predict_files = glob.glob(predict_location + "/*.wav")  # check for wav files
+        if len(all_predict_files) <= 0:
             print(f'Subdirectory {predict_location} has no .h5 or .wav files, exiting ...')
         else:
-            logger.info(f'Found {len(allpfiles)} prediction .wav files.')
-            predwav_mode = True
+            logger.info(f'Found {len(all_predict_files)} prediction .wav files.')
+            predict_wav_mode = True
     else:
-        logger.info(f'Found {len(allpfiles)} prediction .h5 files.')
+        logger.info(f'Found {len(all_predict_files)} prediction .h5 files.')
 
     if len(predict_logfile) == 0:
         logger.info(f'Warning, predict location {predict_location} has no prediction log files.')
@@ -1134,51 +1127,51 @@ def main():
     logger.info(f'Only running specified subset of {len(mixids)} mixtures')
 
     enable_asr_warmup = False
-    if wer_method == 'none':
+    if asr_method == 'none':
         fnb = 'metric_spenh_'
-    elif wer_method == 'google':
+    elif asr_method == 'google':
         fnb = 'metric_spenh_ggl_'
-        logger.info(f'WER enabled with method {wer_method}')
+        logger.info(f'ASR enabled with method {asr_method}')
         enable_asr_warmup = True
-    elif wer_method == 'deepgram':
+    elif asr_method == 'deepgram':
         fnb = 'metric_spenh_dgram_'
-        logger.info(f'WER enabled with method {wer_method}')
+        logger.info(f'ASR enabled with method {asr_method}')
         enable_asr_warmup = True
-    elif wer_method == 'aixplain_whisper':
-        fnb = 'metric_spenh_whspx_' + whisper_model + '_'
-        logger.info(f'WER enabled with method {wer_method} and whisper model {whisper_model}')
+    elif asr_method == 'aixplain_whisper':
+        fnb = 'metric_spenh_whspx_' + asr_model_name + '_'
+        logger.info(f'ASR enabled with method {asr_method} and whisper model {asr_model_name}')
         enable_asr_warmup = True
-    elif wer_method == 'whisper':
-        fnb = 'metric_spenh_whspl_' + whisper_model + '_'
-        logger.info(f'WER enabled with method {wer_method} and whisper model {whisper_model}')
+    elif asr_method == 'whisper':
+        fnb = 'metric_spenh_whspl_' + asr_model_name + '_'
+        logger.info(f'ASR enabled with method {asr_method} and whisper model {asr_model_name}')
         enable_asr_warmup = True
-    elif wer_method == 'aaware_whisper':
-        fnb = 'metric_spenh_whspaaw_' + whisper_model + '_'
-        logger.info(f'WER enabled with method {wer_method} and whisper model {whisper_model}')
+    elif asr_method == 'aaware_whisper':
+        fnb = 'metric_spenh_whspaaw_' + asr_model_name + '_'
+        logger.info(f'ASR enabled with method {asr_method} and whisper model {asr_model_name}')
         enable_asr_warmup = True
-    elif wer_method == 'faster_whisper':
-        fnb = 'metric_spenh_fwhsp_' + whisper_model + '_'
-        logger.info(f'WER enabled with method {wer_method} and whisper model {whisper_model}')
+    elif asr_method == 'faster_whisper':
+        fnb = 'metric_spenh_fwhsp_' + asr_model_name + '_'
+        logger.info(f'ASR enabled with method {asr_method} and whisper model {asr_model_name}')
         enable_asr_warmup = True
     else:
-        logger.error(f'Unrecognized WER method: {wer_method}')
+        logger.error(f'Unrecognized ASR method: {asr_method}')
         return
 
     if enable_asr_warmup:
         DEFAULT_SPEECH = split(DEFAULT_NOISE)[0] + '/speech_ma01_01.wav'
         audio = read_audio(DEFAULT_SPEECH)
         logger.info(f'Warming up asr method, note for cloud service this could take up to a few min ...')
-        asr_chk = calc_asr(audio, engine=wer_method, whisper_model_name=whisper_model)
+        asr_chk = calc_asr(audio, engine=asr_method, whisper_model_name=asr_model_name)
         logger.info(f'Warmup completed, results {asr_chk}')
 
     MP_GLOBAL.mixdb = mixdb
     MP_GLOBAL.predict_location = predict_location
-    MP_GLOBAL.predwav_mode = predwav_mode
+    MP_GLOBAL.predict_wav_mode = predict_wav_mode
     MP_GLOBAL.truth_est_mode = truth_est_mode
     MP_GLOBAL.enable_plot = enable_plot
     MP_GLOBAL.enable_wav = enable_wav
-    MP_GLOBAL.wer_method = wer_method
-    MP_GLOBAL.whisper_model = whisper_model
+    MP_GLOBAL.asr_method = asr_method
+    MP_GLOBAL.asr_model_name = asr_model_name
 
     # Individual mixtures use pandas print, set precision to 2 decimal places
     # pd.set_option('float_format', '{:.2f}'.format)
@@ -1255,7 +1248,7 @@ def main():
             ofname = join(predict_location, fnb + 'summary_truest.txt')
 
         with open(ofname, 'w') as f:
-            print(f'WER enabled with method {wer_method}, whisper model, if used: {whisper_model}', file=f)
+            print(f'ASR enabled with method {asr_method}, whisper model, if used: {asr_model_name}', file=f)
             print(f'Speech enhancement metrics avg over all {len(all_mtab1_sorted_nom99)} non -99 SNR mixtures:',
                   file=f)
             print(all_nom99_mean.to_frame().T.round(2).to_string(float_format=lambda x: "{:.2f}".format(x),
@@ -1318,7 +1311,7 @@ def main():
         label = f'Extraction statistics stats over {num_mix} mixtures:'
         pd.DataFrame([label]).to_csv(csv_name, **header_args)
         all_metrics_table_2.describe().round(2).to_csv(csv_name, **table_args)
-        label = f'WER enabled with method {wer_method}, whisper model, if used: {whisper_model}'
+        label = f'ASR enabled with method {asr_method}, whisper model, if used: {asr_model_name}'
         pd.DataFrame([label]).to_csv(csv_name, **header_args)
 
         if not truth_est_mode: