Code in Python

Speech_To_Text/main.py

@@ -0,0 +1,38 @@
+import whisper
+
+model = whisper.load_model('base')
+
+import gradio as gr # Import gradio as gr
+import time
+
+def transcribe(audio):
+  #time.sleep(3)
+  #step 1: load audio
+  audio = whisper.load_audio(audio)
+  audio = whisper.pad_or_trim(audio)
+
+  #step 2: make a log-Mel Spectrogram and move to the same device as the model
+  mel = whisper.log_mel_spectrogram(audio).to(model.device)
+
+  #step 3: detect the spoken laguage
+  _, probs = model.detect_language(mel)
+  print(f"Detected language: {max(probs, key=probs.get)}")
+
+  #step 4: decode the audio
+  options = whisper.DecodingOptions()
+  result = whisper.decode(model, mel, options)
+
+  #step 5: print the recognized text
+  return result.text
+
+gr.Interface( # Use gr.Interface
+    title = 'Speech to Text using API',
+    fn=transcribe,
+    inputs=[
+        # Removed the 'source' argument and changed 'type' to 'microphone'.
+        gr.Audio(type="filepath")
+    ],
+    outputs=[
+        "textbox"
+    ],
+    live=True).launch()

Speech_To_Text/readme.md

@@ -0,0 +1,24 @@
+A Speech-to-Text (STT) model is an AI-powered system that converts spoken language into written text. These models are widely used in applications like virtual assistants, transcription services, accessibility tools, and voice-controlled devices.
+
+How It Works
+Audio Input: The model takes an audio signal as input, which can be a live speech or a recorded file.
+Preprocessing: The audio is processed to remove noise, normalize volume, and convert it into a suitable format (e.g., spectrogram or Mel-frequency cepstral coefficients - MFCCs).
+Feature Extraction: The model extracts relevant features from the audio, such as phonemes (basic sound units of speech).
+Acoustic Model: It maps the extracted features to corresponding phonetic units.
+Language Model: It refines the raw phonetic transcription by applying linguistic rules, grammar, and context.
+Text Output: The final output is a structured, readable text representation of the spoken words.
+Types of Speech-to-Text Models
+Traditional ASR (Automatic Speech Recognition) – Uses Hidden Markov Models (HMM) and Gaussian Mixture Models (GMM).
+Deep Learning-Based Models – Uses Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), and Transformer-based architectures like Whisper by OpenAI or DeepSpeech by Mozilla.
+End-to-End Models – Uses a single deep learning model (e.g., Transformer, Conformer) to directly map audio to text without separate acoustic and language models.
+Popular Speech-to-Text Models
+Whisper (OpenAI) – Multilingual, robust to noise, works well with various accents.
+DeepSpeech (Mozilla) – Uses deep learning to achieve high accuracy.
+Kaldi – Open-source toolkit for speech recognition.
+Google Speech-to-Text API – Cloud-based solution with high accuracy.
+Vosk – Lightweight and offline-capable STT system.
+Challenges in STT Models
+Background Noise – Can affect transcription accuracy.
+Accents and Dialects – Hard to generalize for all users.
+Code-Switching – Handling multiple languages in one sentence.
+Low-Resource Languages – Less training data available.