srinathlaka/speech-emotion-recognition

Speech Emotion Recognition (CNN, RAVDESS)

A small convolutional network that classifies the emotion in a spoken clip from its mel-spectrogram. Trained on RAVDESS with a speaker-independent split, so the reported accuracy reflects performance on completely unheard voices.

• 🔗 Live demo: https://speech-emotion-recognition-apc1.onrender.com
• 💻 Code: https://github.com/srinathlaka/speech-emotion-recognition

Classes

angry, calm, disgust, fearful, happy, neutral, sad, surprised (indices 0–7, alphabetical).

Results

Metric	Value
Test accuracy (unseen speakers)	46.1%
Best validation accuracy	60.0%
Random baseline	12.5%

Errors are concentrated among emotions of similar arousal (angry↔happy↔surprised, calm↔sad↔neutral). The model is strongest on surprised (~88% recall) and weakest on neutral (underrepresented and acoustically ambiguous).

How the input is built

Each audio clip is converted the same way at train and inference time:

1. Load mono at sr = 22050.
2. Mel-spectrogram with n_mels = 128, then power_to_db(ref=np.max).
3. Pad/truncate the time axis to 130 frames → shape (128, 130).
4. Scale from dB to [0, 1] via (x + 80) / 80.
5. Add batch + channel dims → (1, 1, 128, 130).

Usage (ONNX Runtime)

import numpy as np
import librosa
import onnxruntime as ort

EMOTIONS = ["angry", "calm", "disgust", "fearful",
            "happy", "neutral", "sad", "surprised"]

session = ort.InferenceSession("emotion_cnn.onnx")


def preprocess(path):
    y, _ = librosa.load(path, sr=22050)
    mel = librosa.power_to_db(librosa.feature.melspectrogram(y=y, sr=22050, n_mels=128), ref=np.max)
    if mel.shape[1] > 130:
        mel = mel[:, :130]
    else:
        mel = np.pad(mel, ((0, 0), (0, 130 - mel.shape[1])), mode="minimum")
    mel = (mel + 80.0) / 80.0
    return mel[None, None].astype(np.float32)


logits = session.run(None, {session.get_inputs()[0].name: preprocess("clip.wav")})[0][0]
probs = np.exp(logits) / np.exp(logits).sum()
print(EMOTIONS[int(probs.argmax())], float(probs.max()))

Limitations

• Trained only on RAVDESS (acted North-American English speech); expect lower accuracy on spontaneous speech, other languages, or noisy audio.
• The neutral class is unreliable.
• Intended as a learning/portfolio project, not for real-world decision-making.