HumSpeechBlend Dataset: Humming vs Speech Detection

📌 Overview

HumSpeechBlend is a dataset designed to fine-tune Voice Activity Detection (VAD) models to distinguish between humming and actual speech. Current VAD models often misclassify humming as speech, leading to incorrect segmentation in speech processing tasks. This dataset provides a structured collection of humming audio interspersed with speech to help improve model accuracy.

📜 Dataset Creation Strategy

To build this dataset, the following methodology was used:

1. Humming Audio Collection: Various humming recordings were sourced from the "MLEnd-Hums-and-Whistles" dataset.
2. Speech Insertion: Short speech segments were extracted from "Global Recordings Network" datasets.
3. Mixing Strategy:
   • Speech can be longer or shorter than the humming segment.
   • Speech is randomly inserted at different timestamps in the humming audio.
   • Speech timestamps were carefully annotated to facilitate supervised learning.

🔹 Metadata Explanation

The dataset includes the following metadata columns:

Column Name	Description
file_name	The path to the final mixed audio file (humming + speech).
speech_ts	The timestamps where speech appears within the mixed audio file.
humming_song	The song or source from which the humming was derived.
humming_Interpreter	The individual or source providing the humming. More info in MLEndHWD_Interpreter_Demographics.csv
humming_audio_used	humming audio path in the original dataset.
humming_transcript	Transcription of the humming from whisper-large-v3-turbo.
globalrecordings_audio_used	Speech segment sourced from Global Recordings Network.
globalrecordings_audio_ts_used	The start and end timestamps of the speech segment in the original recording.

📥 Download and Usage

🛠️ Loading the Dataset

Since the dataset does not have predefined splits, you can load it using the following code:

import pandas as pd
from datasets import load_dataset

# Load dataset from Hugging Face
dataset = load_dataset("CuriousMonkey7/HumSpeechBlend", split=None)  # No predefined splits

# Load metadata
metadata = pd.read_feather("metadata.feather")  # Load the Feather metadata
print(metadata.head())

🔊 Loading Audio Files

To work with the audio files:

import torchaudio

waveform, sample_rate = torchaudio.load("data/audio1.wav")
print(f"Sample Rate: {sample_rate}, Waveform Shape: {waveform.shape}")

📄 Citation

If you use this dataset, please cite it accordingly.

@dataset{HumSpeechBlend,
  author = {Sourabh Saini},
  title = {HumSpeechBlend: Humming vs Speech Dataset},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/CuriousMonkey7/HumSpeechBlend}
}