[Chore] update ai model (#40)

ai/example.py

@@ -14,7 +14,7 @@
 
     pro_input = {
         "user-audio": "example1.m4a",
-        "practice-sentence": "It's autumn now, and the leaves are turning beautiful colors.",
+        "practice-sentece": "It's autumn now, and the leaves are turning beautiful colors.",
         "tip": "Say 'aw-tum,' not 'ay-tum.'"
     }
 
@@ -23,16 +23,16 @@
     pro_feedback = ft.ProFeedback(pro_input)
 
     ## Sample Output ##
-    print("** Feedback for Contextual Task: ", com_feedback)
-    # {'positive-feedback': 'You are very creative! I like your imagination.', 
+    print("** Feedback for Communication Task: ", com_feedback)
+    # {'positive-feedback': 'You are very creative! I like your imagination.',
     #  'negative-feedback': "Let's try to describe what we see in the picture. First, look at the sky. What colors can you see there?",
     #   'enhanced-answer': 'In the sky, I can see yellow, orange, pink, and blue.'}
 
     print("** Feedback for Pronunciation Task: ", pro_feedback)
-    # {'transcription': 'ITS AUTUMN NOW AND THE LEAVES ARE TURNING BEAUTIFUL COLORS', 
-    #   'wrong_idx': {'minor': [2, 9], 'major': []}, 
-    #   'pronunciation_score': 0.7, 
-    #   'decibel': 46.90759735625882, 
-    #   'speech_rate': 2.347417840375587, 
-    #   'positive-feedback': 'Pronunciation is correct. Keep up the good work!', 
+    # {'transcription': 'ITS AUTUMN NOW AND THE LEAVES ARE TURNING BEAUTIFUL COLORS',
+    #   'wrong_idx': {'minor': [2, 9], 'major': []},
+    #   'pronunciation_score': 0.7,
+    #   'decibel': 46.90759735625882,
+    #   'speech_rate': 2.347417840375587,
+    #   'positive-feedback': 'Pronunciation is correct. Keep up the good work!',
     #   'negative-feedback': ' '}

ai/fluentify.py

@@ -4,15 +4,13 @@
 import librosa
 import numpy as np
 import stable_whisper
-import torch
 import vertexai
 import yaml
 from evaluate import load
-from transformers import AutoTokenizer, AutoFeatureExtractor, AutoModelForCTC
 from vertexai.preview.generative_models import GenerativeModel, Part
 
 from utils.utils import text2dict
-from utils.word_process import get_incorrect_idx
+from utils.word_process import get_incorrect_idxes
 
 
 # gcloud auth application-default login
@@ -25,21 +23,19 @@ def __init__(self):
         self.lang_model = GenerativeModel("gemini-pro")
         self.current_path = os.path.dirname(__file__)
         self.gcs_path = "gs://fluentify-412312.appspot.com"
-        self.shared_path = "./shared-data"
-        self.ars_whisper_model = stable_whisper.load_model('base')
+        self.ars_model = stable_whisper.load_model('base')
 
         with open(os.path.join(self.current_path, 'data/prompt.yaml'), 'r', encoding='UTF-8') as file:
             self.prompt = yaml.load(file, Loader=yaml.FullLoader)
 
-        self.audio_path = self.shared_path + "/audio"
-        self.tokenizer = AutoTokenizer.from_pretrained("facebook/wav2vec2-base-960h")
-        self.ars_w2v_model = AutoModelForCTC.from_pretrained("facebook/wav2vec2-base-960h")
-        self.feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base-960h",
-                                                                      sampling_rate=16000)
+        self.audio_path = "./data/audio"
+        # self.tokenizer =  AutoTokenizer.from_pretrained("facebook/wav2vec2-base-960h")
+        # self.ars_w2v_model = AutoModelForCTC.from_pretrained("facebook/wav2vec2-base-960h")
+        # self.feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base-960h", sampling_rate=16000)
         self.wer = load("wer")
 
-        self.speech_rate_threshold_h = 2.5
-        self.speech_rate_threshold_l = 1.0
+        self.speech_rate_threshold_h = 4.5
+        self.speech_rate_threshold_l = 2.5
         self.decibel_threshold_h = 95
         self.decibel_threshold_l = 45
 
@@ -49,37 +45,26 @@ def __init__(self):
     def GetWer(self, transcription, ground_truth):
         # WER score is from 0 to 1
         wer_score = self.wer.compute(predictions=[transcription], references=[ground_truth])
-        incorrect_idx = get_incorrect_idx(ground_truth, transcription)
-        incorrect_idx = {"minor": incorrect_idx["substitutions"], "major": incorrect_idx["deletions"]}
-        return wer_score, incorrect_idx
-
-    def ASR(self, audio_path):
-        ## TODO : Audio File Path ####
-        input_audio = librosa.load(audio_path, sr=22000)[0]
-        input_values = self.feature_extractor(input_audio, return_tensors="pt", padding="longest").input_values
-        rms = librosa.feature.rms(y=input_audio)
+        incorrect_idxes = get_incorrect_idxes(ground_truth, transcription)
+        self.wer_score = wer_score
+        self.incorrect_idxes = incorrect_idxes["substitutions"] + incorrect_idxes["deletions"]
 
-        with torch.no_grad():
-            logits = self.ars_w2v_model(input_values).logits[0]
-            pred_ids = torch.argmax(logits, axis=-1)
-        outputs = self.tokenizer.decode(pred_ids, output_word_offsets=True)
-        time_offset = self.ars_w2v_model.config.inputs_to_logits_ratio / self.feature_extractor.sampling_rate
-
-        transcription = outputs[0]
-        word_offsets = [
-            {
-                "word": d["word"],
-                "start_time": round(d["start_offset"] * time_offset, 2),
-                "end_time": round(d["end_offset"] * time_offset, 2),
-                "time": round(d["end_offset"] * time_offset - d["start_offset"] * time_offset, 2),
-            }
-            for d in outputs.word_offsets
-        ]
-
-        time = word_offsets[-1]["end_time"]
-        speech_rate = len(word_offsets) / word_offsets[-1]["end_time"]
+    def GetDecibel(self, audio_path):
+        input_audio, _ = librosa.load(audio_path, sr=22000)
+        rms = librosa.feature.rms(y=input_audio)
         decibel = 20 * math.log10(np.mean(rms) / 0.00002)
-        return transcription, time, speech_rate, decibel
+        return decibel
+
+    def EvaluatePro(self, audio_path):
+        transcription = self.ars_model.transcribe(audio_path)
+        self.pro_transcription = transcription.text
+
+        word_offset = transcription.segments
+        start, end = word_offset[0].start, word_offset[-1].end
+        duration = end - start
+
+        self.speech_rate = len(self.pro_transcription.split(' ')) / duration
+        self.decibel = self.GetDecibel(audio_path)
 
     def Score2Des(self, output):
         feedback = {
@@ -113,11 +98,12 @@ def ScoreMap(self, input, threshold):
             return 5
 
     def ComFeedback(self, input):
-        image = Part.from_uri(f"{self.gcs_path}/{input['img']}", mime_type="image/jpeg")
-        response = self.ars_whisper_model.transcribe(f"{self.audio_path}/{input['user-audio']}")
+        image = Part.from_uri(f"{self.gcs_path}/img/{input['img']}", mime_type="image/jpeg")
+        response = self.ars_model.transcribe(f"{self.audio_path}/{input['user-audio']}")
+        self.com_trnascription = response.text
+        output = {"transcription": self.com_trnascription}
 
-        output = {"transcription": response.text}
-        #### TODO : Feedback focusing more on sentence formulation ##### 
+        #### TODO : Feedback focusing more on sentence formulation #####
         prompt = f"{self.prompt['con-feedback']}".format(context=input["context"], question=input["question"],
                                                          answer=input["answer"], user_answer=output['transcription'])
         response = self.multimodal_model.generate_content([prompt, image])
@@ -131,30 +117,29 @@ def ComFeedback(self, input):
             return None
 
     def ProFeedback(self, input):
-        ground_truth = input["practice-sentence"]
-        transcription, time, speech_rate, decibel = self.ASR(f"{self.audio_path}/{input['user-audio']}")
-        wer_score, incorrect_idx = self.GetWer(transcription.upper(), ground_truth.upper())
-
+        ground_truth = input["practice-sentece"]
+        self.EvaluatePro(f"{self.audio_path}/{input['user-audio']}")
+        self.GetWer(self.pro_transcription.upper(), ground_truth.upper())
         sentence_lst = ground_truth.split(" ")
 
-        pronunciation_score = self.ScoreMap(1 - wer_score, 1)
-        speed_socre = self.ScoreMap(speech_rate, self.speech_rate_threshold)
-        volume_score = self.ScoreMap(decibel, self.decibel_threshold)
+        pronunciation_score = self.ScoreMap(1 - self.wer_score, 1)
+        speed_score = self.ScoreMap(self.speech_rate, self.speech_rate_threshold)
+        volume_score = self.ScoreMap(self.decibel, self.decibel_threshold)
 
         output = {
-            "transcription": transcription,
-            "incorrect_indexes": incorrect_idx['major'],
+            "transcription": self.pro_transcription,
+            "incorrect_indexes": self.incorrect_idxes,
+            "decibel": self.decibel,
+            "speech_rate": self.speech_rate,
             "pronunciation_score": pronunciation_score,  # higher the better
-            "voulume_score": volume_score,  # higher the better
-            "speed_socre": speed_socre,
-            "decibel": decibel,
-            "speech_rate": speech_rate
+            "volume_score": volume_score,  # higher the better
+            "speed_score": speed_score
         }
 
         feedback = self.Score2Des(output)
 
-        ## TODO : Decide whether to use only major errors or only minor errors ## 
-        # only if there is major incorrect pronunciation 
+        ## TODO : Decide whether to use only major errors or only minor errors ##
+        # only if there is major incorrect pronunciation
         if len(output["incorrect_indexes"]) > 0:
             for idx in output["incorrect_indexes"]:
                 feedback['incorrect_pronunciation'] = f"Your pronunciation for {sentence_lst[idx]} is not correct"

ai/talking-video-gen.py

@@ -1,15 +1,14 @@
 # Imports the Google Cloud stt library
 import json
 import os
-import time
-from datetime import datetime
 
 import cv2
 import pandas as pd
 import stable_whisper
 from PIL import Image
 from google.cloud import texttospeech
 from moviepy.editor import VideoFileClip, AudioFileClip
+from tqdm import tqdm
 
 
 # gcloud auth application-default login
@@ -26,16 +25,17 @@ def __init__(self):
         self.data_path = './data'
         self.character_path = f'{self.data_path}/character'
 
-        self.ts = datetime.now().strftime("%m%d%H%M%S")
-        self.dir_path = f'{self.data_path}/video_output/{self.ts}'
+        # self.ts  = datetime.now().strftime("%m%d%H%M%S")
+        # self.dir_path = f'{self.data_path}/video_output/{self.ts}'
+        self.dir_path = f'{self.data_path}/pro-video-output'
 
         self.audio_path = self.dir_path + '/output.mp3'
         self.video_path = self.dir_path + "/output.mp4"
         self.wordoffset_path = self.dir_path + '/word_offset.json'
         self.output_path = self.dir_path + '/fianal-output.mp4'
 
-        if not os.path.exists(self.dir_path):
-            os.makedirs(self.dir_path)
+        # if not os.path.exists(self.dir_path):
+        #     os.makedirs(self.dir_path)
 
         img_lst = os.listdir(self.character_path)
         self.close_img_lst = [img for img in img_lst if 'close' in img]
@@ -96,7 +96,7 @@ def VideoGen(self):
         open_img = self.close_img_lst[0]
 
         for word in self.wordoffset.iloc:
-            ## CLOSE ## 
+            ## CLOSE ##
             if word['content'] == None:
                 # close_img = random.choice(self.close_img_lst)
                 for _ in range(int(word['time'])):
@@ -122,12 +122,14 @@ def VideoGen(self):
         video.release()
 
     def MergeAudioVideo(self):
+
         video_clip = VideoFileClip(self.video_path)
         audio_clip = AudioFileClip(self.audio_path)
         final_clip = video_clip.set_audio(audio_clip)
         final_clip.write_videofile(self.output_path)
 
-    def Text2TalkingVideo(self, sentence):
+    def Text2TalkingVideo(self, sentence, fname="feedback"):
+        self.output_path = f'{self.dir_path}/{fname}.mp4'
         self.Text2Speech(sentence)
         self.Speech2Text()
         self.WordOffset()
@@ -138,9 +140,37 @@ def Text2TalkingVideo(self, sentence):
 
 if __name__ == "__main__":
     tg = TalkingGenerator()
-    sentence = "The most important thing to keep in mind when designing this car would be to make sure that it is made of a metal that the superhero can control with his mind."
+
+    with open('./data/pro-data.json') as f:
+        pro_dataset = json.load(f)
+
+    with open('./data/com-data.json') as f:
+        com_dataset = json.load(f)
+
+        for i, pro_data in tqdm(enumerate(pro_dataset)):
+            pro_data.update({'practice-sentence-w-tip': pro_data['practice-sentence'] + ' ' + pro_data['tip']})
+            sentence = pro_data["practice-sentence-w-tip"]
+            fname = pro_data["id"]
+            tg.Text2TalkingVideo(sentence, fname)
+
+    # for i,com_data in tqdm(enumerate(com_dataset)):
+    #     sentence = com_data["question"]
+    #     fname = com_data["id"]
+    #     tg.Text2TalkingVideo(sentence, fname)
+
+    with open('./data/pro-data.json', 'w') as f:
+        json.dump(pro_dataset, f, indent=4)
+
+    # with open('./data/com-data.json', 'w') as f:
+    #     json.dump(com_dataset, f, indent=4)
+    # for com_data in tqdm(com_dataset):
+    #     sentence = com_data['question']
+    #     fname = com_data['img'].split('.')[0]
+    #     tg.Text2TalkingVideo(sentence, fname)
+
+    # sentence ="The most important thing to keep in mind when designing this car would be to make sure that it is made of a metal that the superhero can control with his mind."
     #  sentence =  "Let's imagine that you are a brave captain of a big ship. You are sailing on the high seas. Suddenly, you see a beautiful sunset. Look at this picture and tell me..."
-    time_ = time.time()
-    tg.Text2TalkingVideo(sentence)
-    time_ = time.time() - time_
-    print(time_)
+    # time_  = time.time()
+    # tg.Text2TalkingVideo(sentence=,fname=)
+    # time_  = time.time() - time_
+    # print(time_)

ai/utils/utils.py

@@ -1,5 +1,5 @@
-import json
 import ast
+import json
 
 
 def text2dict(text):
@@ -10,4 +10,48 @@ def text2dict(text):
     try:
         return ast.literal_eval(text)
     except:
-        return json.loads(text) if text else None
+        try:
+            return json.loads(text)
+        except:
+            text = text.split('positive-feedback')[-1]
+            if 'enhanced-answer' in text:
+                pos, neg_w_enh = text.split('negative-feedback')
+                neg, enh = neg_w_enh.split('enhanced-answer')
+                pos, neg, enh = clean(pos), clean(neg), clean(enh)
+                return {"positive-feedback": pos, "negative-feedback": neg, "enhanced-answer": enh}
+            else:
+                pos, neg = text.split('negative-feedback')
+                pos, neg = clean(pos), clean(neg)
+                return {"positive-feedback": pos, "negative-feedback": neg}
+
+
+def clean(text):
+    text = text.strip()
+    if text.startswith('"') or text.startswith("'"):
+        text = text[1:]
+    if text.endswith('"') or text.endswith("'"):
+        text = text[:-1]
+
+    if ":" in text:
+        text = text.replace(":", "")
+    if "," in text:
+        text = text.replace(",", "")
+    if "{" in text:
+        text = text.replace("{", "")
+    if "}" in text:
+        text = text.replace("}", "")
+
+    if "\'s" in text:
+        text = text.replace("\'s", " 's")
+    if "\'m" in text:
+        text = text.replace("\'s", " 'm")
+
+    text = text.strip()
+
+    if text.startswith('"') or text.startswith("'"):
+        text = text[1:]
+    if text.endswith('"') or text.endswith("'"):
+        text = text[:-1]
+
+    text = text.strip()
+    return text