--- language: - en license: unknown task_categories: - automatic-speech-recognition - audio-classification - text-classification annotations_creators: - expert-generated source_datasets: - earnings22 multilinguality: - monolingual size_categories: - 1K A prediction, or fragment thereof, that has no phonetic correspondence with the content of the audio signal analyzed by the ASR model. Annotators marked words or phrases using the following taxonomy: ### Hallucination A word or phrase that does not correspond to any spoken content in the audio recording. ### Looping An erroneous repetition of a word or phrase that is present in the speech signal. ### Hallucination Looping An erroneous repetition of content that is itself hallucinated and does not correspond to the audio signal. Each audio file was annotated independently by two annotators, with disagreements resolved by a third annotator acting as an arbitrator. In addition to hallucination annotation, all reference transcripts were manually reviewed and corrected when necessary. During this process, approximately 14% of examples were identified as partially inaudible or ambiguous. The resulting dataset contains: - Human-annotated hallucination spans - Human-annotated looping spans - Human-annotated hallucination-looping spans - Corrected reference transcripts - Utterance-level hallucination labels - Predictions from multiple ASR systems Importantly, HALAS was intentionally constructed to contain a high proportion of hallucinations and therefore **does not represent the real-world prevalence of ASR hallucinations**. However, all hallucinations originate from genuine speech recordings rather than synthetic corruption or non-speech inputs. --- ## Supported Tasks - Automatic Speech Recognition (ASR) - Hallucination Detection - Hallucination Mitigation - Error Detection - Speech Recognition Evaluation - Speech Processing Research --- ## Languages - English The source recordings originate from earnings-call recordings featuring speakers from 27 countries. --- ## License HALAS is derived from the Earnings22 dataset and augments it with human-generated hallucination annotations and corrected reference transcripts. The original Earnings22 dataset is publicly distributed under the **Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)** license. Users of HALAS must comply with the terms of the Earnings22 license in addition to any conditions specified for this dataset. The full license text is available at: https://creativecommons.org/licenses/by-sa/4.0/ HALAS contains original human annotations produced by the authors. Unless otherwise specified in the official repository, these annotations are intended to be distributed under the same license as the source dataset for compatibility with the underlying Earnings22 data. **Note:** Users should consult the official HALAS repository for the authoritative licensing terms of the released version. --- ## Dataset Sources ### Repository https://github.com/DSP-AGH/HALAS/tree/main ### Paper **HALAS: A Human-Annotated Dataset of Hallucinations of Modern ASR Systems** Mateusz Barański, Jan Jasiński, Julitta Bartolewska, Marcin Witkowski, and Konrad Kowalczyk. Accepted at Interspeech 2026. --- ## Data Fields The dataset contains the following fields: | Field | Description | |---------|-------------| | `audio_id` | Unique identifier of the audio segment from Earnings22 | | `audio_duration` | Duration of the audio segment in seconds | | `e22_reference_text` | Original reference transcript provided by Earnings22 | | `corrected_reference_text` | Human-corrected reference transcript | | `model_prediction` | ASR prediction generated by the corresponding model | | `model_label` | Utterance-level hallucination label indicating whether hallucination was present | | `model_hallucination_text` | Text span annotated as hallucination | | `model_hallucination_json` | Structured span annotations containing hallucination locations and categories | | `split` | Dataset split (`train` or `test`) | The `model` field corresponds to one of the following ASR systems: - `whisper_large_v2` - `whisper_large_v3` - `whisper_large_v3_turbo` - `crisper_whisper` - `canary` - `canary_flash` - `parakeet` - `phi4`* - `granite`* Models marked with an asterisk (*) are outside the scope of related paper, as they were trained on the Earnings22 dataset. --- ## Annotation Methodology ### Annotators HALAS was annotated by **10 professional annotators** with English proficiency at B2 level or higher. ### Annotation Procedure For each audio sample: 1. The audio recording was reviewed. 2. Audio quality issues were identified (e.g., non-English speech, overlapping speakers, inaudible segments). 3. Annotators reviewed ASR predictions while listening to the audio. 4. Hallucinated spans were marked using the HALAS taxonomy. 5. A second annotator independently repeated the process. 6. A third annotator resolved disagreements and finalized annotations. ### Quality Control Each sample was reviewed by two independent annotators. Agreement between annotators was high: - Cohen's κ = 0.87 This level of agreement indicates strong consistency in hallucination identification. Additionally, all reference transcripts were revalidated and corrected by the arbitrator. --- ## Data Splits | Split | Files | |---------|---------| | Train | 2,866 | | Test | 745 | The train/test partition was created by: - Splitting by source meeting - Stratifying according to: - Average WER - Hallucination rate - Audio duration To ensure a stable benchmark, the test set includes only recordings: - Longer than 1 second - Containing at least 3 words Dataset statistics: | Split | Hallucination Rate | |---------|---------| | Train | 33.6% | | Test | 22.6% | --- ## Dataset Characteristics ### Source Audio The source recordings originate from the Earnings22 dataset: - Approximately 119 hours of audio - Earnings-call recordings - Speakers from 27 countries - Real-world recording conditions Candidate segments were selected from portions exhibiting high disagreement between ASR systems. Audio files can be downloaded using following script: ``` import csv from datasets import load_dataset import soundfile as sf import os # Load dataset filtered_dataset = load_dataset("distil-whisper/earnings22", 'chunked', split='test') # Create output directory output_dir = "e22" os.makedirs(output_dir, exist_ok=True) # CSV file setup csv_file = os.path.join(output_dir, "metadata.csv") with open(csv_file, mode='w', newline='', encoding='utf-8') as file: writer = csv.writer(file) writer.writerow(["filepath", "text"]) # Write header for audio in filtered_dataset: seg_id = audio["segment_id"] audio_array = audio["audio"]["array"] sample_rate = audio["audio"]["sampling_rate"] name = audio["file_id"] name = "_".join([seg_id, name]) file_path = os.path.join(output_dir, name + ".wav") text = audio['transcription'] # Save audio file sf.write(file_path, audio_array, sample_rate) # Write metadata to CSV writer.writerow([file_path, text]) print(f"Saved: {file_path}") print("All files and metadata saved successfully.") ``` --- ## Bias, Risks, and Limitations HALAS intentionally oversamples difficult audio segments. As a consequence: 1. The hallucination frequency does not reflect real-world deployment conditions. 2. The dataset is biased toward audio examples that produce disagreement among ASR systems. 3. Results obtained on HALAS should not be interpreted as estimates of hallucination prevalence in production systems. 4. The dataset focuses exclusively on English earnings-call recordings and may not generalize to other domains, languages, or acoustic environments. 5. Some audio samples contain challenging conditions such as accents, low-quality recordings, or partially inaudible speech. 6. Although references were manually corrected, certain recordings remain inherently ambiguous due to poor audio quality. Researchers should therefore view HALAS primarily as a benchmark for hallucination detection and analysis rather than a representative sample of everyday ASR usage. --- ## Recommendations Researchers using HALAS should: 1. Report whether evaluations are performed at utterance level or span level. 2. Account for the intentionally elevated hallucination rate when interpreting results. 3. Distinguish hallucinations from ordinary transcription errors. 4. Consider both structural and semantic evaluation metrics. 5. Evaluate generalization on additional datasets whenever possible. Because HALAS contains naturally occurring hallucinations from real speech recordings, it provides a substantially more realistic benchmark than datasets constructed from non-speech or artificially corrupted audio. --- ## Citation ### BibTeX ```bibtex @inproceedings{baranski2026halas, title={HALAS: A Human-Annotated Dataset of Hallucinations of Modern ASR Systems}, author={Barański, Mateusz and Jasiński, Jan and Bartolewska, Julitta and Witkowski, Marcin and Kowalczyk, Konrad}, booktitle={Proceedings of Interspeech 2026}, year={2026} } ``` ### APA Barański, M., Jasiński, J., Bartolewska, J., Witkowski, M., & Kowalczyk, K. (2026). *HALAS: A Human-Annotated Dataset of Hallucinations of Modern ASR Systems*. In *Proceedings of Interspeech 2026*. --- ## Dataset Card Authors Mateusz Barański, Jan Jasiński, Julitta Bartolewska, Marcin Witkowski, and Konrad Kowalczyk. Signal Processing Group Institute of Electronics AGH University of Krakow, Poland --- ## Dataset Card Contact - Mateusz Barański — mbaranski@agh.edu.pl - Jan Jasiński — jjasinsk@agh.edu.pl Alternatively, please use the issue tracker in the official repository once available.