ARTPARK-IISc/Vaani-Atypical-Speech-Corpus

Project Euphonia is a public initiative led by Google that aims to improve Automatic Speech Recognition (ASR) for individuals with atypical speech. To date, most of Project Euphonia’s work has focused on English, resulting in outcomes such as the Android application Project Relate, which generates personalized speech recognition models in English. In recent years, the project has expanded its data collection efforts to additional languages, including French, Spanish, Japanese, and Hindi.

The Vaani Atypical Speech Corpus is a collaboration between Project Vaani and Project Euphonia to collect atypical speech data in Indic languages. This partnership leverages Vaani’s extensive expertise and nationwide reach in India alongside Project Euphonia’s experience in gathering speech data from historically underserved communities. The dataset is intended to support research toward improving ASR systems for such speakers, with potential applications in voice dictation, real-time communication and accurate closed captioning.

---

In the data collection process, once we identify the targeted speakers, as we followed in Vaani, we will present images and ask participants to describe or speak about them. Each participant will provide recordings for 20 images, with each recording lasting approximately 20–40 seconds. The data is collected with the assistance of Karya and due to the challenges in understanding atypical speech, audio segments are transcribed by individuals familiar with the participant’s speech patterns (e.g., family members or educators), to ensure they capture the speaker’s intended meaning as fully as possible, rather than a strict verbatim wording. While the nature of the disability of each speaker is not collected, the speech disabilities represented in this dataset include learning disabilities (mild–high), autism, cerebral palsy, Down syndrome, speech and hearing impairments, and intellectual and developmental disabilities.

Collected data undergoes rigorous automated and manual validation by language experts to ensure high-quality standards. Although not specialists in atypical speech, reviewers assess every audio segment for technical integrity while verifying that the speaking style is natural and voluntary. Simultaneously, they ensure content is safe, private, and relevant to the accompanying visual context.

---

The dataset is currently intended for research use only.

Dataset Summary

Language	Segments	Speakers	Total Duration	Mean Duration
Hindi	688	25	296.6 min	25.5 s
Marathi	290	8	194.5 min	39.7 s
Telugu	189	7	109.1 min	34.6 s
Total	1,167	40	600.2 min	—

Subsets / Configurations

Config	Description
Hindi	Hindi segments only
Marathi	Marathi segments only
Telugu	Telugu segments only
all	All languages combined (default)

Prompt media (images and videos used to elicit speech) are stored in data/image/ (1,093 images) and data/video/ (84 videos). They are not audio configs — filter by the prompt_file_type metadata field instead.

Usage

from datasets import load_dataset
from huggingface_hub import hf_hub_download
from PIL import Image
import requests

REPO = "ARTPARK-IISc/Vaani-Atypical-Speech-Corpus"
access_token="Your access token"

# Load a specific language
ds_hindi   = load_dataset(REPO, "Hindi",token=access_token)
ds_marathi = load_dataset(REPO, "Marathi",token=access_token)
ds_telugu  = load_dataset(REPO, "Telugu",token=access_token)

# Load all languages (default)
ds = load_dataset(REPO,token=access_token)


# Filter by prompt type
image_segs = ds["train"].filter(lambda x: x["prompt_file_type"] == "image")
video_segs = ds["train"].filter(lambda x: x["prompt_file_type"] == "video")

Associate prompt image / video with each audio segment

Each row has a prompt_file_name field (e.g. Vaani-Euphonia-Pilot-Image-42.jpg) that points to the corresponding prompt file stored in data/image/ or data/video/.

from huggingface_hub import hf_hub_download
from PIL import Image

REPO = "ARTPARK-IISc/Vaani-Euphonia"

def get_prompt_file(row):
    """Download and return the prompt image or video path for a given row."""
    folder = "data/image" if row["prompt_file_type"] == "image" else "data/video"
    local_path = hf_hub_download(
        repo_id=REPO,
        repo_type="dataset",
        filename=f"{folder}/{row['prompt_file_name']}",
    )
    return local_path

# Example: iterate and pair audio with its prompt image
for row in ds["train"].select(range(5)):
    audio      = row["audio"]          # {"array": ..., "sampling_rate": 16000, "path": ...}
    prompt     = get_prompt_file(row)  # local path to image/video file

    if row["prompt_file_type"] == "image":
        img = Image.open(prompt)
        img.show()

    print(f"Speaker : {row['speaker_id']}")
    print(f"Language: {row['language']}")
    print(f"Prompt  : {row['prompt_file_name']}  ({row['prompt_file_type']})")
    print(f"Text    : {row['transcription']}")
    print()

Metadata Fields

Field	Description
audio	Audio object (WAV, 16 kHz mono)
speaker_id	Unique contributor identifier
language	Hindi / Marathi / Telugu
prompt_file_type	Prompt media type: image or video
prompt_file_name	Prompt file used to elicit the speech
transcription	Intended transcription
difficulty_level	Low / Medium / High (from QC understandability rating)
verbatim_transcription	Word-for-word record by a non-specialist transcriber
duration_s	Audio duration in seconds