Added audio file trimming into worker preprocessing

modules/server/master_server.py

@@ -45,6 +45,7 @@ class Application(tornado.web.Application):
             (r"/worker/ws/speech", WorkerWebSocketHandler)
         ]
         tornado.web.Application.__init__(self, handlers, **settings)
+        self.connected_worker = 0
         self.available_workers = set()
         self.waiting_client = set()
         self.num_requests_processed = 0
@@ -61,7 +62,7 @@ class Application(tornado.web.Application):
 
     def display_server_status(self):
         logging.info('#'*50)
-        logging.info("Available workers: %s" % str(len(self.available_workers)))
+        logging.info("Connected workers: %s (Available: %s)" % (str(self.connected_worker),str(len(self.available_workers))))
         logging.info("Waiting clients: %s" % str(len(self.waiting_client)))
         logging.info("Requests processed: %s" % str(self.num_requests_processed))
             
@@ -158,6 +159,7 @@ class WorkerWebSocketHandler(tornado.websocket.WebSocketHandler):
     def open(self):
         self.client_handler = None 
         self.application.available_workers.add(self)
+        self.application.connected_worker += 1
         self.application.check_waiting_clients()
         logging.debug("Worker connected")
         self.application.display_server_status()
@@ -169,7 +171,6 @@ class WorkerWebSocketHandler(tornado.websocket.WebSocketHandler):
             logging.debug("Message received from worker:" + message)
         else:  
             if 'transcription' in json_msg.keys(): #Receive the file path to process
-                print(json_msg['transcription'])
                 logging.debug("Response send by worker : %s" % json.dumps({'transcript':json_msg['transcription'].encode('utf-8')}))
                 self.client_handler.receive_response(json.dumps({'transcript':json_msg['transcription']}))
                 self.client_handler = None
@@ -187,6 +188,7 @@ class WorkerWebSocketHandler(tornado.websocket.WebSocketHandler):
             self.client_handler.send_error("Worker closed")
         logging.debug("WORKER WebSocket closed")
         self.application.available_workers.discard(self)
+        self.application.connected_worker -= 1
         self.application.display_server_status()
 
 def main():

modules/worker_offline/requirements.txt

 ws4py
 configparser
-tenacity 
\ No newline at end of file
+tenacity
+pydub
\ No newline at end of file

modules/worker_offline/scripts/decode.sh

@@ -31,7 +31,7 @@ sysRootName=$(echo $(basename $sysdir)|cut -f1 -d"=")
 #        java -Xmx2024m -jar $lvcsrRootDir/tools/lium_spkdiarization-8.4.1.jar  \
 #            --fInputMask=$file --sOutputMask=$datadir/$fileRootName.seg --doCEClustering $fileRootName
 	duration=`soxi -D $file`
-	echo "$fileRootName 1 0 $duration M S U S0" | sort -nk3 > $datadir/$fileRootName.seg
+#	echo "$fileRootName 1 0 $duration M S U S0" | sort -nk3 > $datadir/$fileRootName.seg
     fi
     if [ $stage -le 2 ]; then
         # Generate kaldi input for offline decoding
@@ -50,7 +50,7 @@ sysRootName=$(echo $(basename $sysdir)|cut -f1 -d"=")
 		    echo "$tag sox $file -t wav -r 16000 -c 1 - |"
 		done) > $datadir/wav.scp
 		cat $datadir/wav.scp | awk '{ print $1, $1, "A"; }' > $datadir/reco2file_and_channel
-		echo validate_data_dir.sh
+#		echo validate_data_dir.sh
 		$lvcsrRootDir/scripts/utils/validate_data_dir.sh --no-text --no-feats $datadir
 		$lvcsrRootDir/scripts/utils/fix_data_dir.sh $datadir
 	fi
@@ -67,7 +67,7 @@ sysRootName=$(echo $(basename $sysdir)|cut -f1 -d"=")
 
 	if [ $stage -le 4 ]; then
         if [ ! -f $transdir/trans.1 ]; then
-            echo "run fmllr decoding"
+ #           echo "run fmllr decoding"
             $lvcsrRootDir/scripts/steps/decode.sh --nj $decode_nj --cmd "$decode_cmd" --num-threads $num_threads --skip-scoring "true" \
             $gmmdir/Graph $datadir $transdir || exit 1
         fi
@@ -80,5 +80,5 @@ sysRootName=$(echo $(basename $sysdir)|cut -f1 -d"=")
 	### Rescoring with LM
 	### Get CTM and STM files
 
-echo "End...."
+#echo "End...."
 

modules/worker_offline/signal_trimming.py

+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Thu Jan 18 17:32:23 2018
+
+@author: rbaraglia
+"""
+
+from pydub import AudioSegment
+
+
+def detect_leading_silence(sound, silence_threshold=-50.0, chunk_size=100):
+    '''
+    sound is a pydub.AudioSegment
+    silence_threshold in dB
+    chunk_size in ms
+
+    iterate over chunks until you find the first one with sound
+    '''
+    trim_ms = 0 # ms
+
+    assert chunk_size > 0 # to avoid infinite loop
+    while sound[trim_ms:trim_ms+chunk_size].dBFS < silence_threshold and trim_ms < len(sound):
+        trim_ms += chunk_size
+    return trim_ms
+
+def average_power_level(sound, chunk_size=100):
+    trim_ms = 0 # ms
+    nb_chunk = 0
+    avg_power = 0.0
+    assert chunk_size >0
+    while trim_ms < len(sound):
+        trim_ms += chunk_size
+        if (sound[trim_ms:trim_ms+chunk_size].dBFS != -float('Inf')):
+            avg_power += sound[trim_ms:trim_ms+chunk_size].dBFS
+            nb_chunk += 1
+    avg_power = avg_power/nb_chunk
+    return avg_power
+
+    '''
+    trim_silence_segments remove silence (or background noise) from an audio wav file.
+    It working by trimming signal at the beginning and the end that is below the overall power level
+    input_file is a .wav file path
+    output_file is a .wav file path
+    chunk_size in ms
+    threshold_factor between 0 and 1
+    side_effect_accomodation is a number of chunk that will be kept at the beginning and end despite being below the threshold
+    
+    '''
+def trim_silence_segments(input_file,output_file, chunk_size=100, threshold_factor=0.85, side_effect_accomodation=0):
+    #sound = AudioSegment.from_file("/home/rbaraglia/data/SG/audio-18_01_18/rec---2018-01-18_081957.wav", format="wav")
+    sound = AudioSegment.from_file(input_file, format="wav")
+    avg_power = average_power_level(sound)
+    start_trim = detect_leading_silence(sound,silence_threshold= threshold_factor * avg_power)
+    end_trim = detect_leading_silence(sound.reverse(), silence_threshold= threshold_factor * avg_power)
+    
+    duration = len(sound)
+    trimmed_sound = sound[start_trim if start_trim - chunk_size*side_effect_accomodation < 0 else start_trim - chunk_size*side_effect_accomodation : duration-end_trim if end_trim + chunk_size*side_effect_accomodation > duration else duration-end_trim + chunk_size*side_effect_accomodation]
+    trimmed_sound.export(output_file, format="wav")
+    
+

modules/worker_offline/worker.cfg

@@ -6,4 +6,4 @@ server_target : /worker/ws/speech
 [worker_params]
 temp_file_location : wavs/
 decoder_command : "./scripts/decode.sh systems/models"
-preprocessing : false
\ No newline at end of file
+preprocessing : true
\ No newline at end of file

modules/worker_offline/worker_offline.py

@@ -14,6 +14,7 @@ import subprocess
 import configparser
 import re
 import tenacity
+from signal_trimming import *
 
 from ws4py.client.threadedclient import WebSocketClient
 
@@ -50,13 +51,14 @@ class WorkerWebSocket(WebSocketClient):
                 self.client_uuid = json_msg['uuid']
                 self.fileName = self.client_uuid.replace('-', '')
                 self.file = json_msg['file'].decode('base64')
-
-                with open('./wavs/'+self.fileName+'.wav', 'wb') as f:
+                self.filepath = TEMP_FILE_PATH+self.fileName+'.wav'
+                with open(self.filepath, 'wb') as f:
                     f.write(self.file)
                 logging.debug("FileName received: %s" % self.fileName)
                 # TODO: preprocessing ? (sox python)
                 if PREPROCESSING:
-                    pass
+                    logging.debug("Trimming signal")
+                    trim_silence_segments(self.filepath,self.filepath, chunk_size=100, threshold_factor=0.85, side_effect_accomodation=0)
                 # Offline decoder call
                 
                 logging.debug(DECODER_COMMAND + ' ' + TEMP_FILE_PATH + self.fileName+'.wav')