densities

app.py

@@ -78,7 +78,7 @@ def plott(g1,w1,l1,s1,g2,w2,l2,s2):
 
 
 
-bl = gr.Blocks(theme=gr.themes.Glass())
+bl = gr.Blocks()#theme=gr.themes.Glass())
 
 with bl:
 	
@@ -108,8 +108,8 @@ with bl:
 					#### Select data (2)
 					"""
 						)
-					gmenu2 = gr.Dropdown(choices=['[NONE]'] + grouplist,label="Group", value='[NONE]')
-					wmenu2 = gr.Dropdown(label="Word", choices=['[NONE]'])
+					gmenu2 = gr.Dropdown(choices=['[NONE]'] + grouplist,label="Group", value='A:L')
+					wmenu2 = gr.Dropdown(label="Word", choices=['[ALL]'] + [n for n,v in worddict['A:L']])
 					lmenu2 = gr.Radio(choices=["L1", "L2","All"],label="Speaker group",value="L1")
 					smenu2 = gr.Dropdown(["Annotated", "MFA"],label="Source",value="Annotated")
 
@@ -118,7 +118,7 @@ with bl:
 
 	
 			btn = gr.Button(value="Update Plot")
-			plo = gr.Plot()
+			plo = gr.Plot(value=plott('AL:','[ALL]',"L1","Annotated",'A:L','[ALL]',"L1","Annotated"))
 			btn.click(plott, [gmenu1,wmenu1,lmenu1,smenu1,gmenu2,wmenu2,lmenu2,smenu2], plo)
 
 
@@ -140,6 +140,21 @@ with bl:
 			"""
 				 )
 
+			
+			gr.Markdown(
+			"""
+			### About
+
+			This annotated data and its demo application accompany the paper 
+			*Assessed and Annotated Vowel Lengths in Spoken Icelandic Sentences\
+			for L1 and L2 Speakers: A Resource for Pronunciation Training*, \
+			Caitlin Laura Richter, Kolbrún Friðriksdóttir, Kormákur Logi Bergsson, \
+			Erik Anders Maher, Ragnheiður María Benediktsdóttir, Jon Gudnason - NoDaLiDa/Baltic-HLT 2025, Tallinn, Estonia.
+			
+
+			"""
+				 )
+			
 			gr.Markdown(
 				""" 
 				## Demo: Viewing the data
@@ -148,6 +163,9 @@ with bl:
 				Available speaker groups are native Icelandic speakers (L1), second-language speakers (L2), or all. 
 				Data source options are gold (human) annotations or automated Montreal Forced Aligner (MFA).
 
+				The display is a scatter plot of vowel and consonant durations,
+				supplemented with density plots for each dimension separately.
+
 				The general expectation is that, all else being equal, syllables with long stressed vowels 
 				followed by short consonants have a higher vowel:(vowel+consonant) duration ratio, 
 				while syllables with short stressed vowels followed by long consonants have a lower ratio. 
@@ -161,7 +179,6 @@ with bl:
 				)
 
 
-			
 			gr.Markdown(
 				""" 
 			## Accessing the data
@@ -171,25 +188,69 @@ with bl:
 			or [tsv](https://github.com/catiR/length-contrast-data-isl/blob/main/Data/Length_in_spoken_icelandic.tsv) files. 
 			See [the paper](https://github.com/catiR/length-contrast-data-isl/blob/main/Data/133_Annotated_Vowel_Lengths.pdf) 
 			for complete information. 
+			"""
+				 )
 
+			gr.Markdown(
+				""" 
 			Audio is available from [Clarin](https://repository.clarin.is/repository/xmlui/) (Samrómur). 
 			The 'collection' field plus recording filename in the annotations metadata 
 			specify the original audio file, including which Samrómur collection it is found in. 
 			"""
 				 )
 
-
 			gr.Markdown(
-			"""
-			### About
-
-			This annotated data and its demo application accompany the paper 
-			*Assessed and Annotated Vowel Lengths in Spoken Icelandic Sentences\
-			for L1 and L2 Speakers: A Resource for Pronunciation Training*, \
-			Caitlin Laura Richter, Kolbrún Friðriksdóttir, Kormákur Logi Bergsson, \
-			Erik Anders Maher, Ragnheiður María Benediktsdóttir, Jon Gudnason - NoDaLiDa/Baltic-HLT 2025, Tallinn, Estonia.
+				""" 
+Annotation records are in the following scheme: 
+
+```
+[ { recording: source-file-id.wav,
+    collection: samromur-collection,
+    speaker_lang: L1/L2,
+    word: target-word,
+    word_context: { 
+        normalised: normalised-carrier-sentence-text,
+        before: sentence-context-preceding-token,
+        after: sentence-context-following-token 
+    },
+    gold_annotation: {
+        target_word_start: seconds,
+        target_word_end: seconds,
+        prevowel: [ { 
+                phone: ipa-character,
+                start: seconds,
+                end: seconds, 
+                }, 
+                { phone2 ... } ,
+        ], 
+        vowel: [ { 
+                phone: ipa-character,
+                start: seconds,
+                end: seconds, 
+                }, 
+        ], 
+        postvowel: [ { 
+                phone: ipa-character,
+                start: seconds,
+                end: seconds, 
+                }, 
+        ]
+    },
+    mfa_annotation : {
+     ... as for gold ...
+    }
+ },
+]
+```	
+		
+				"""
+				 )
 			
 
+
+
+			gr.Markdown(
+			"""
 			### Contact caitlinr@ru.is about bugs, feedback, or collaboration!
 
 			"""

vowel_length.py

@@ -5,7 +5,9 @@ import pandas as pd
 import matplotlib
 matplotlib.use('Agg')
 import matplotlib.pyplot as plt
-
+from scipy.stats import gaussian_kde
+#from scipy.spatial import KDTree
+#from sklearn.neighbors import NearestNeighbors
 
 # make subsets of words for convenience
 def make_sets(db,shorts,longs):
@@ -70,7 +72,7 @@ def get_tk_data(tk,shorts,longs):
 		tot_start, tot_end = plist[0]['start'],plist[-1]['end']
 		tot_dur = tot_end-tot_start
 		return tot_dur
-
+	
 	tkdat = {}
 	tkdat['word'] = tk['word']
 	tkdat['speaker_lang'] = tk['speaker_lang']
@@ -89,7 +91,7 @@ def get_tk_data(tk,shorts,longs):
 		tkdat[f'{s}_post_dur'] = _merge_intervals(tk[f'{s}_annotation']['postvowel'])
 		tkdat[f'{s}_word_dur'] = tk[f'{s}_annotation']['target_word_end'] -\
 		  tk[f'{s}_annotation']['target_word_start']
-
+	
 	return tkdat
 
 
@@ -118,18 +120,44 @@ def setup(annot_json):
 				'mömmu', 'nærri', 'palla', 'raggi', 'skeggi', 'snemma', 'sunna',
 				'tommi', 'veggi','vinnur', 'ásta'])
 
+	# very basic remove about 5 outliers > 350ms
+	cut=0.35
+
 	with open(annot_json, 'r') as handle:
 		db = json.load(handle)
 
 	sets = make_sets(db,shorts,longs)
 
 	db = [get_tk_data(tk,shorts,longs) for tk in db]
+	db = [t for t in db if ((t['gold_v_dur'] <=cut) and (t['gold_post_dur'] <=cut))]
 	dat = pd.DataFrame.from_records(db)
 	dat = prep_dat(dat)
 
 	return sets,dat
 
 
+def kldiv(s1,s2):
+	_log = lambda x: np.log2(x) if x != 0 else 0
+	_log = np.vectorize(_log)
+	
+	n, m = len(s1), len(s2)
+	d = s1.shape[1]
+	assert d == 2 == s2.shape[1]
+
+	k = 1
+	while True:
+		knn1 = NearestNeighbors(n_neighbors = k+1).fit(s1)
+		nnDist1 = knn1.kneighbors(s1)[0][:, k]
+		if not nnDist1.all():
+			k += 1
+		else:
+			break
+	knn2 = NearestNeighbors(n_neighbors = k).fit(s2)
+	nnDist2 = knn2.kneighbors(s1)[0][:, k-1]
+	kl = (d/n) * sum(_log(nnDist2/nnDist1)) + _log((m/(n-1)))
+	return kl
+
+		
 
 def vgraph(dat1,l1,src1,lab1,dat2,l2,src2,lab2):
 
@@ -161,6 +189,7 @@ def vgraph(dat1,l1,src1,lab1,dat2,l2,src2,lab2):
 
 		return vdurs, cdurs, rto, cc
 
+	plt.close()
 	
 	vd1,cd1,ra1,cl1 = _gprep(dat1,l1,src1)
 	lab1 += f'\n Ratio: {ra1:.3f}'
@@ -171,12 +200,16 @@ def vgraph(dat1,l1,src1,lab1,dat2,l2,src2,lab2):
 
 
 	fig, ax = plt.subplots(figsize=(9,7))
-	ax.set_xlim(0.0,350)
-	ax.set_ylim(0.0,350)
+	#ax.set_xlim(0.0, 350)
+	#ax.set_ylim(0.0, 350)
 
 	ax.scatter(vd1,cd1,marker = mk1, label = lab1,
 				   c = [cl1 + (.7,)], edgecolors = [cl1] )
 
+	marginals = [(vd1, 'x', l1, cl1),
+				(cd1, 'y', l1, cl1)]
+	
+	#kld = None
 	if lab2:
 		vd2,cd2,ra2,cl2 = _gprep(dat2,l2,src2)
 		lab2 += f'\n Ratio: {ra2:.3f}'
@@ -186,22 +219,58 @@ def vgraph(dat1,l1,src1,lab1,dat2,l2,src2,lab2):
 			mk2 = '>'
 		ax.scatter(vd2,cd2, marker = mk2, label = lab2,
 					c = [cl2 + (.05,)], edgecolors = [cl2] )
+		#s1 = np.transpose(np.array([vd1,cd1]))
+		#s2 = np.transpose(np.array([vd2,cd2]))
+		#klda = kldiv(s1,s2)
+		#if klda:
+		#	kldb = kldiv(s2,s1)
+		#	kldsym = np.mean([klda,kldb])
+		#	if not np.isnan(kldsym):
+		#		ax.scatter([-300],[-300],c = 'white',label = f'\nKLDiv: {kldsym:.2f}')
+			
+		marginals += [(vd2, 'x', l2, cl2),
+					(cd2, 'y', l2, cl2)]
+
+	#fig.legend(loc=8,ncols=2)
+	leg = fig.legend(loc=7,frameon=False)
+	for t in leg.get_texts():
+		t.set_verticalalignment("center_baseline")
 
+	ax.axline((0,0),slope=1,color="darkgray")
 
-	ax.set_title("Stressed vowel & following consonant(s) duration" )
+	marginals = [m for m in marginals if len(m[0])>9]
+	lsts = {'L1': 'solid' , 'L2': 'dashed' , 'All': 'dashdot'}
+	for values, axt, lng, lcl in marginals:
+		kde = gaussian_kde(values, bw_method='scott')
+		pts = np.linspace(np.min(values), np.max(values))
+		dens = kde.pdf(pts)
+		scf=2500
+		lst = lsts[lng]
+		#l2dat = ax.plot(pts, [350-(scf*i) for i in dens], linestyle=lst, color = lcl)
+		l2dat = ax.plot(pts, [350+(scf*i) for i in dens], linestyle=lst, color = lcl, clip_on=False)
+		if axt == 'y':
+			for l2d in l2dat:
+				xln = l2d.get_xdata()
+				yln = l2d.get_ydata()
+				l2d.set_xdata(yln)
+				l2d.set_ydata(xln)
+				fig.canvas.draw()
+				#ax.draw_artist(l2d)
+
+				
+	ax.set_xlim(0.0, 350)
+	ax.set_ylim(0.0, 350)
+
+	ax.set_title("Stressed vowel & following consonant(s) duration" , fontsize=16, y=-.155)
 	ax.set_xlabel("Vowel duration (ms)")
 	ax.set_ylabel("Consonant duration (ms)")
-	#fig.legend(loc=8,ncols=2)
-	fig.legend(loc=7)
-	
-	ax.axline((0,0),slope=1,color="darkgray")
-	
+
 	fig.tight_layout()
-	#fig.subplots_adjust(bottom=0.15)
-	fig.subplots_adjust(right=0.75)
+	fig.subplots_adjust(bottom=0.13)
+	fig.subplots_adjust(right=0.72)
 
 	#plt.xticks(ticks=[50,100,150,200,250,300],labels=[])
 	#plt.yticks(ticks=[100,200,300],labels=[])
-
+	
 	return fig