Add Streamlit UI for real-time SER predictions; update README and requirements

[Saurabh620]

Motivation

• Provide an interactive Streamlit interface to run inference with the trained Speech Emotion Recognition (SER) model for quick predictions.
• Improve repository documentation and usage instructions to include running the Streamlit app and model path expectations.
• Ensure the project dependencies include Streamlit so the UI can be executed reproducibly.

Description

• Add a new streamlit_app.py that loads a Keras/TensorFlow model, extracts MFCC features from uploaded audio, runs inference, and displays predicted emotion, class probabilities, and a bar chart.
• Implement extract_mfcc_from_bytes to read audio bytes with soundfile, resample to 22050 Hz, slice/pad a 3s window, compute 40 MFCCs, and shape the result as (1, 40, 1).
• Use st.cache_resource to cache the loaded model via get_model, and include robust model-loading fallback between keras and tensorflow.keras.
• Update README.md with step-by-step run instructions and a new "Streamlit UI (optimized)" section, and add streamlit==1.45.1 to requirements.txt.

Testing

• No automated tests were run for these changes.

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 329129b476

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Normalize MFCC features with training scaler

main() feeds raw MFCC vectors directly into model.predict, but the training notebook standardizes inputs (MainProgram.ipynb builds X_normalized with StandardScaler and trains on that tensor). Skipping the same scaler at inference creates a train/serve feature mismatch that can materially distort class probabilities and predicted emotions for normal user uploads; the app should load and apply the exact scaler used during training before prediction.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Remove unsupported uploaded audio extensions

The uploader advertises mp3 and m4a, but decoding is done solely through soundfile.read in extract_mfcc_from_bytes with no fallback or error handling. In environments using the pinned soundfile==0.10.3.post1, these compressed formats are often not decodable (especially m4a), so selecting such a file can raise a runtime decode error instead of producing a prediction; either restrict allowed types to guaranteed codecs or add a robust fallback decoder path.

Useful? React with 👍 / 👎.