scripts/clusterprosody.py

import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import soundfile as sf
from collections import defaultdict
from dtw import dtw
from sklearn_extra.cluster import KMedoids
from scipy import stats
from copy import deepcopy
import os, librosa, json


# based on original implementation by 
# https://colab.research.google.com/drive/1RApnJEocx3-mqdQC2h5SH8vucDkSlQYt?authuser=1#scrollTo=410ecd91fa29bc73
# by magnús freyr morthens 2023 supported by rannís nsn



def z_score(x, mean, std):
    return (x - mean) / std


# given a sentence and list of its speakers + their alignment files,
# return a dictionary of word alignments
def get_word_aligns(norm_sent, aln_paths):
    """
    Returns a dictionary of word alignments for a given sentence.
    """
    word_aligns = defaultdict(list)
    slist = norm_sent.split(" ")
    
    for spk,aln_path in aln_paths:
        with open(aln_path) as f:
            lines = f.read().splitlines()
        lines = [l.split('\t') for l in lines]
        try:
            assert len(lines) == len(slist)
            word_aligns[spk] = [(w,float(s),float(e)) for w,s,e in lines]
        except:
            print(slist, lines, "<---- something didn't match")
    return word_aligns
             
    

def get_pitches(start_time, end_time, fpath):
    """
    Returns an array of pitch values for a given speech.
    Reads from .f0 file of Time, F0, IsVoiced
    """
    
    with open(fpath) as f:
        lines = f.read().splitlines()
        lines = [[float(x) for x in line.split()] for line in lines]    # split lines into floats
        pitches = []
        

        # find the mean of all pitches in the whole sentence
        mean = np.mean([line[1] for line in lines if line[2] == 1])
        # find the std of all pitches in the whole sentence
        std = np.std([line[1] for line in lines if line[2] == 1])

        low = min([p for t,p,v in lines if v == 1]) - 1

        for line in lines:
            time, pitch, is_pitch = line
            
            if start_time <= time <= end_time:
                if is_pitch == 1:
                    pitches.append(z_score(pitch, mean, std))
                else:
                    pitches.append(z_score(low, mean, std))
                    #pitches.append(-0.99)
    
    return pitches
    
    

# jcheng used energy from esps get_f0
# get f0 says (?) :
#The RMS value of each record is computed based on a 30 msec hanning
#window with its left edge placed 5 msec before the beginning of the
#frame.
# jcheng z-scored the energys, per file.
# TODO: implement that. ?
# not sure librosa provides hamming window in rms function directly
# TODO handle audio that not originally .wav
def get_rmse(start_time, end_time, wpath, znorm = True):
    """
    Returns an array of RMSE values for a given speech.
    """
    
    audio, sr = librosa.load(wpath, sr=16000)
    hop = 80
    #segment = audio[int(np.floor(start_time * sr)):int(np.ceil(end_time * sr))]
    rmse = librosa.feature.rms(y=audio,frame_length=480,hop_length=hop)
    rmse = rmse[0] 
    if znorm:
        rmse = stats.zscore(rmse)
    segment = rmse[int(np.floor(start_time * sr/hop)):int(np.ceil(end_time * sr/hop))]
    #idx = np.round(np.linspace(0, len(rmse) - 1, pitch_len)).astype(int)
    return segment#[idx]


# may be unnecessary depending how rmse and pitch window/hop are calculated already
def downsample_rmse2pitch(rmse,pitch_len):
    idx = np.round(np.linspace(0, len(rmse) - 1, pitch_len)).astype(int)
    return rmse[idx]

    

# parse user input string to usable word indices for the sentence
# TODO handle more user input cases
def parse_word_indices(start_end_word_index):
    ixs = start_end_word_index.split('-')
    if len(ixs) == 1:
        s = int(ixs[0])
        e = int(ixs[0])
    else:
        s = int(ixs[0])
        e = int(ixs[-1])
    return s-1,e-1


# take any (1stword, lastword) or (word) 
#   unit and prepare data for that unit
def get_data(norm_sent,path_key,start_end_word_index):
    """
    Returns a dictionary of pitch, rmse, and spectral centroids values for a given sentence/word combinations.
    """
    
    s_ix, e_ix = parse_word_indices(start_end_word_index)
    words = '_'.join(norm_sent.split(' ')[s_ix:e_ix+1])
    
    align_paths = [(spk,pdict['aln']) for spk,pdict in path_key]
    word_aligns = get_word_aligns(norm_sent, align_paths)
    
    data = defaultdict(list)
    align_data = defaultdict(list)
    playable_audio = {}

    for spk, pdict in path_key:
        word_al = word_aligns[spk]
        start_time = word_al[s_ix][1]
        end_time = word_al[e_ix][2]

        seg_aligns =  word_al[s_ix:e_ix+1]
        seg_aligns = [(w,round(s-start_time,2),round(e-start_time,2)) for w,s,e in seg_aligns]
                    
        pitches = get_pitches(start_time, end_time, pdict['f0'])
        
        rmses = get_rmse(start_time, end_time, pdict['wav'])
        rmses = downsample_rmse2pitch(rmses,len(pitches))
        #spectral_centroids = get_spectral_centroids(start_time, end_time, id, wav_dir, len(pitches))
            
        pitches_cpy = np.array(deepcopy(pitches))
        rmses_cpy = np.array(deepcopy(rmses))
        d = [[p, r] for p, r in zip(pitches_cpy, rmses_cpy)]
        #words = "-".join(word_combs)
        data[f"{words}**{spk}"] = d
        align_data[f"{words}**{spk}"] = seg_aligns
        playable_audio[spk] = (pdict['wav'], start_time, end_time)
                
    return words, data, align_data, playable_audio

    

def dtw_distance(x, y):
    """
    Returns the DTW distance between two pitch sequences.
    """
    
    alignment = dtw(x, y, keep_internals=True)
    return alignment.normalizedDistance
    
    
    
# recs is a sorted list of rec IDs
# all recs/data contain the same words
# rec1 and rec2 can be the same
def pair_dists(data,words,recs):

    dtw_dists = []
    
    for rec1 in recs:
        key1 = f'{words}**{rec1}'
        val1 = data[key1]
        for rec2 in recs:
            key2 = f'{words}**{rec2}'
            val2 = data[key2]
            dtw_dists.append((f"{rec1}**{rec2}", dtw_distance(val1, val2)))
            
    return dtw_dists



# TODO 
# make n_clusters a param with default 3
def kmedoids_clustering(X):
    kmedoids = KMedoids(n_clusters=3, random_state=0).fit(X)
    y_km = kmedoids.labels_
    return y_km, kmedoids



def match_tts(clusters, speech_data, tts_data, tts_align, words, seg_aligns, voice):

    
    tts_info = []
    for label in set([c for r,c in clusters]):
        recs = [r for r,c in clusters if c==label]
        dists = []
        for rec in recs: 
            key = f'{words}**{rec}'
            dists.append(dtw_distance(tts_data, speech_data[key]))
        tts_info.append((label,np.nanmean(dists)))
        
    tts_info = sorted(tts_info,key = lambda x: x[1])
    best_cluster = tts_info[0][0]
    best_cluster_score = tts_info[0][1]
    
    matched_data = {f'{words}**{r}': speech_data[f'{words}**{r}'] for r,c in clusters if c==best_cluster}

    # now do graphs of matched_data with tts_data
    # and report best_cluster_score
    
    mid_cluster = tts_info[1][0]
    mid_data = {f'{words}**{r}': speech_data[f'{words}**{r}'] for r,c in clusters if c==mid_cluster}
    bad_cluster = tts_info[2][0]
    bad_data = {f'{words}**{r}': speech_data[f'{words}**{r}'] for r,c in clusters if c==bad_cluster}
    
    #tts_fig_p = plot_pitch_tts(matched_data,tts_data, tts_align, words,seg_aligns,best_cluster,voice)
    tts_fig_p, best_cc = plot_one_cluster(words,'pitch',matched_data,seg_aligns,best_cluster,tts_data=tts_data,tts_align=tts_align,voice=voice)
    fig_mid_p, mid_cc = plot_one_cluster(words,'pitch',mid_data,seg_aligns,mid_cluster)
    fig_bad_p, bad_cc = plot_one_cluster(words,'pitch',bad_data,seg_aligns,bad_cluster)

    
    tts_fig_e, _ = plot_one_cluster(words,'rmse',matched_data,seg_aligns,best_cluster,tts_data=tts_data,tts_align=tts_align,voice=voice)
    fig_mid_e, _ = plot_one_cluster(words,'rmse',mid_data,seg_aligns,mid_cluster)
    fig_bad_e, _ = plot_one_cluster(words,'rmse',bad_data,seg_aligns,bad_cluster)
    
    
    # TODO
    # not necessarily here, bc paths to audio files.
    spk_cc_map = [('Best',best_cluster,best_cc), ('Mid',mid_cluster,mid_cc), ('Last',bad_cluster,bad_cc)]
    #playable = audio_htmls(spk_cc_map)
   
    return best_cluster_score, tts_fig_p, fig_mid_p, fig_bad_p, tts_fig_e, fig_mid_e, fig_bad_e
    


def gp(d,s,x):
    return os.path.join(d, f'{s}.{x}')

def gen_tts_paths(tdir,voices):
    plist = [(v, {'wav': gp(tdir,v,'wav'), 'aln': gp(tdir,v,'tsv'), 'f0': gp(tdir,v,'f0')}) for v in voices]
    return plist
    
def gen_h_paths(wdir,adir,f0dir,spks):
    plist = [(s, {'wav': gp(wdir,s,'wav'), 'aln': gp(adir,s,'tsv'), 'f0': gp(f0dir,s,'f0')}) for s in spks]
    return plist
    

# since clustering strictly operates on X,
#  once reduce a duration metric down to pair-distances,
# it no longer matters that duration and pitch/energy had different dimensionality
# TODO option to dtw on 3 feats pitch/ener/dur separately
# check if possible cluster with 3dim distance mat? 
# or can it not take that input in multidimensional space
#  then the 3 dists can still be averaged to flatten, if appropriately scaled

def cluster(norm_sent,orig_sent,h_spk_ids, h_align_dir, h_f0_dir, h_wav_dir, tts_sent_dir, voices, start_end_word_index):

    h_spk_ids = sorted(h_spk_ids)
    nsents = len(h_spk_ids)

    h_all_paths = gen_h_paths(h_wav_dir,h_align_dir,h_f0_dir,h_spk_ids)

    words, h_data, h_seg_aligns, h_playable = get_data(norm_sent,h_all_paths,start_end_word_index)

    dtw_dists = pair_dists(h_data,words,h_spk_ids) 
    
    kmedoids_cluster_dists = []
            
    X = [d[1] for d in dtw_dists]
    X = [X[i:i+nsents] for i in range(0, len(X), nsents)]
    X = np.array(X)
    
    y_km, kmedoids = kmedoids_clustering(X)
    #plot_clusters(X, y_km, words)
    #c1, c2, c3 = [X[np.where(kmedoids.labels_ == i)] for i in range(3)]

    result = zip(X, kmedoids.labels_)
    groups = [[r,c] for r,c in zip(h_spk_ids,kmedoids.labels_)]

    
    tts_all_paths = gen_tts_paths(tts_sent_dir, voices)
    _, tts_data, tts_seg_aligns, tts_playable_segment = get_data(norm_sent,tts_all_paths,start_end_word_index)
    
    for v in voices:
        voice_data = tts_data[f"{words}**{v}"]
        voice_align = tts_seg_aligns[f"{words}**{v}"]
               
        #tts_data, tts_align = get_one_tts_data(tts_sent_dir,v,norm_sent,start_end_word_index)
       
    # match the data with a cluster -----
        best_cluster_score, tts_fig_p, fig_mid_p, fig_bad_p, tts_fig_e, fig_mid_e, fig_bad_e = match_tts(groups, h_data, voice_data, voice_align, words, h_seg_aligns,v)


    audio_html = clusters_audio(groups,h_playable)

    # only supports one voice at a time currently
    return best_cluster_score, tts_fig_p, fig_mid_p, fig_bad_p,  tts_fig_e, fig_mid_e, fig_bad_e, audio_html
    #return words, kmedoids_cluster_dists, group




# generate html panel to play audios for each human cluster
# audios is dict {recording_id : (wav_path, seg_start_time, seg_end_time)}
def clusters_audio(clusters,audios):
    
    html = '''<html><body>'''
    
    for label in set([c for r,c in clusters]):
        recs = [r for r,c in clusters if c==label]
        
        html += '<div>'
        html += f'<h2>Cluster {label}</h2>'
        
        html += '<div>'
        html += '<table><tbody>'
        
        for rec in recs:
            html += f'<tr><td><audio controls id="{rec}">'   #width="20%"> 
            html += f'<source src="{audios[rec][0]}#t={audios[rec][1]*60:.2f},{audios[rec][2]*60:.2f}" type="audio/wav">' 
            html += '</audio></td>'
            html += f'<td>{rec}</td></tr>'
            
        html += '</tbody></table>'
        html += '</div>'
        #html += '<div style="height:2%;background:#e7fefc"></div>'
        
        html += '</div>'
    html += '</body></html>'
    
    return html



# find offsets to visually align start of each word for speakers in cluster
def reset_cluster_times(words,cluster_speakers,human_aligns,tts_align):
    words = words.split('_')

    retimes = [(words[0], 0.0)]
    for i in range(len(words)-1):
        gaps = [human_aligns[spk][i+1][1]-human_aligns[spk][i][1] for spk in cluster_speakers]
        if tts_align:
            gaps.append(tts_align[i+1][1] - tts_align[i][1])
        retimes.append((words[i+1],retimes[i][1]+max(gaps)))
    return retimes

# apply offsets for a speaker
def retime_speaker_xvals(retimes, speaker_aligns, speaker_xvals):
    new_xvals = []
    def xlim(x,i,retimes,speaker_aligns):
        return (x < speaker_aligns[i+1][1]) if i+1<len(retimes) else True
        
    for i in range(len(retimes)):
        wd,st = retimes[i]
        w,s,e = speaker_aligns[i]
        xdiff = st-s
        new_xvals += [x+xdiff for x in speaker_xvals if (x>= s) and xlim(x,i,retimes,speaker_aligns) ]
    
    return [round(x,3) for x in new_xvals]


# interpolate NAN to break lines
def retime_xs_feats(retimes, speaker_aligns, speaker_xvals, feats):
    feat_xvals = retime_speaker_xvals(retimes, speaker_aligns, speaker_xvals)
    xf0 = list(zip(feat_xvals, feats))
    xf = [xf0[0]]
    for x,f in xf0[1:]:
        lx = xf[-1][0]
        if x - lx >= 0.01:
            xf.append((lx+0.005,np.nan))
        xf.append((x,f))
    return [x for x,f in xf], [f for x,f in xf]



def plot_one_cluster(words,feature,speech_data,seg_aligns,cluster_id,tts_data=None,tts_align=None,voice=None):
#(speech_data, tts_data, tts_align, words, seg_aligns, cluster_id, voice):
    colors = ["red", "green", "blue", "orange", "purple", "pink", "brown", "gray", "cyan"]
    cc = 0
    spk_ccs = [] # for external display
    fig = plt.figure(figsize=(10, 5))
    
    if feature.lower() in ['pitch','f0']:
        fname = 'Pitch'
        ffunc = lambda x: [p for p,e in x]
        pfunc = plt.scatter
    elif feature.lower() in ['energy', 'rmse']:
        fname = 'Energy'
        ffunc = lambda x: [e for p,e in x]
        pfunc = plt.plot
    else:
        print('problem with the figure')
        return fig, []
    
    
    # boundary for start of each word
    retimes = reset_cluster_times(words,list(speech_data.keys()),seg_aligns,tts_align)
    if len(retimes)>1:
        for w,bound_line in retimes:
            plt.axvline(x=bound_line, color="gray", linestyle='--', linewidth=1, label=f'Start "{w}"')
    
    plt.title(f"{words} - {fname} - Cluster {cluster_id}")
    
    for k,v in speech_data.items():

        spk = k.split('**')[1]
        word_times = seg_aligns[k]
        
        feats = ffunc(v)
        # datapoint interval is 0.005 seconds
        feat_xvals = [x*0.005 for x in range(len(feats))]

        feat_xvals, feats = retime_xs_feats(retimes,word_times,feat_xvals,feats)
        pfunc(feat_xvals, feats, color=colors[cc], label=f"Speaker {spk}")
        
        #feat_xvals = retime_speaker_xvals(retimes, word_times, feat_xvals)
        #for w, st in reversed(retimes):
        #    w_xvals = [x for x in feat_xvals if x>= st]
        #    w_feats = feats[-(len(w_xvals)):]
        #    pfunc(w_xvals, w_feats, color=colors[cc])
        #    feat_xvals = feat_xvals[:-(len(w_xvals))]
        #    feats = feats[:-(len(w_xvals))]


        spk_ccs.append((spk,colors[cc]))        
        cc += 1
        if cc >= len(colors):
            cc=0

    if voice:
        tfeats = ffunc(tts_data)
        t_xvals = [x*0.005 for x in range(len(tfeats))]
        
        t_xvals, tfeats = retime_xs_feats(retimes, tts_align, t_xvals, tfeats)
        pfunc(t_xvals, tfeats, color="black", label=f"TTS {voice}")
        
        #t_xvals = retime_speaker_xvals(retimes, tts_align, t_xvals)
        #for w, st in reversed(retimes):
        #    tw_xvals = [x for x in t_xvals if x>= st]
        #    tw_feats = tfeats[-(len(tw_xvals)):]
        #    pfunc(tw_xvals, tw_feats, color="black")
        #    t_xvals = t_xvals[:-(len(tw_xvals))]
        #    tfeats = tfeats[:-(len(tw_xvals))]

    #plt.legend()
    #plt.show()
            

    return fig, spk_ccs