Playlogue with SOT Format for Multi-Speaker ASR

Processed version of the Playlogue dataset with Serialized Output Training (SOT) format for multi-speaker automatic speech recognition.

Dataset Description

This dataset contains 158 conversations (~33 hours) of adult-child interactions with:

• Speaker attribution (adult vs child)
• Timestamps for each utterance
• SOT format for end-to-end multi-speaker ASR training
• Audio playback enabled in HuggingFace viewer

Dataset Statistics

Split	Samples	Duration
Train	97	~19 hours
Validation	27	~6 hours
Test	34	~8 hours
Total	158	~33 hours

Corpora Included

• EllisWeismer: Children with/without language delays (play sessions)
• Gleason: Parent-child play interactions
• VanHouten: Examiner-child free play
• Cameron: African American English + Standard American English narratives

Data Format

Each sample contains:

{
    'audio': Audio(sampling_rate=16000),
    'file_id': str,
    'corpus': str,  # EllisWeismer, Gleason, VanHouten, Cameron
    'participant': str,
    'split': str,  # train, val, test
    'duration': float,
    'num_utterances': int,
    'num_child_utterances': int,
    'num_adult_utterances': int,
    'sot_text': str,  # SOT format with speaker tags
    'utterances_json': str  # JSON string with detailed utterances
}

SOT Format

Serialized Output Training (SOT) format embeds speaker information directly in the transcript:

<|startoftranscript|> <|en|> <|transcribe|> 
<|0.7|> <adult> alright <|2.1|> 
<|2.7|> <child> how this go out <|3.4|> 
<|4.1|> <adult> um there we go <|6.0|>
<|endoftranscript|>

Format: <|time|> <speaker> text <|time|>

Usage

Load Dataset

from datasets import load_dataset

# Load dataset
dataset = load_dataset("sulaimank/playlogue")

# Access splits
train = dataset['train']
val = dataset['validation']
test = dataset['test']

# Get sample
sample = train[0]
audio_array = sample['audio']['array']
sampling_rate = sample['audio']['sampling_rate']
sot_text = sample['sot_text']

Filter by Corpus

# Filter EllisWeismer only
ellisweismer = dataset.filter(lambda x: x['corpus'] == 'EllisWeismer')

# Filter Gleason only
gleason = dataset.filter(lambda x: x['corpus'] == 'Gleason')

Filter by Participant

# Get all conversations for participant 'andy'
andy_data = dataset.filter(lambda x: x['participant'] == 'andy')

Training Example

from transformers import WhisperProcessor, WhisperForConditionalGeneration

# Load model
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Extend vocabulary for speaker tokens
new_tokens = ["<adult>", "<child>"]
processor.tokenizer.add_tokens(new_tokens)
model.resize_token_embeddings(len(processor.tokenizer))

# Prepare data
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_features"] = processor.feature_extractor(
        audio["array"], 
        sampling_rate=audio["sampling_rate"]
    ).input_features[0]
    batch["labels"] = processor.tokenizer(batch["sot_text"]).input_ids
    return batch

# Train on SOT format
dataset = dataset.map(prepare_dataset, remove_columns=dataset.column_names)

Citation

Original Playlogue dataset:

@article{kalanadhabhatta2024playlogue,
  title={Playlogue: Dataset and Benchmarks for Analyzing Adult-Child Conversations During Play},
  author={Kalanadhabhatta, Manasa and Rastikerdar, Mohammad Mehdi and Rahman, Tauhidur and Grabell, Adam S. and Ganesan, Deepak},
  journal={Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies},
  volume={8},
  number={4},
  year={2024},
  publisher={ACM},
  url={https://doi.org/10.1145/3699775}
}

License

This dataset follows the same license as the original Playlogue dataset. Users must follow CHILDES ground rules for data usage.

Acknowledgments

• Original Playlogue dataset by Kalanadhabhatta et al.
• CHILDES/TalkBank for source data
• SOT format inspired by multi-speaker ASR research