Bridge2AI-Voice Pediatric Dataset

@article{PhysioNet-b2ai-voice-pediatric-2.0.0,
  author = {Bensoussan, Yael and Sigaras, Alexandros and Rameau, Anais and Elemento, Olivier and Powell, Maria and Dorr, David and Payne, Philip and Ravitsky, Vardit and Bélisle-Pipon, Jean-Christophe and Bahr, Ruth and Watts, Stephanie and Bolser, Donald and Siu, Jennifer and Lerner-Ellis, Jordan and Rudzicz, Frank and Boyer, Micah and Abdel-Aty, Yassmeen and {Ahmed Syed}, Toufeeq and Anibal, James and Aradi, Stephen and Armosh, Kirollos and Martinez, Ana Sophia and Awan, Shaheen and Bedrick, Steven and Beltran, Helena and Bernier, Alexander and Berrios, Moroni and Bevers, Isaac and Blatter, Alden and Brito, Rahul and Brown, Amy and Brown, Johnathan and Casalino, Selina and Costello, John and Dalal, Abhijeet and {De Santiago}, Iris and Diaz-Ocampo, Enrique and Doerksen, Emily and Ebraheem, Mohamed and Eiseman, Ellie and Elmahdy, Mahmoud and Evangelista, Emily and Fletcher, Kenneth and Gallois, Hortense and Garrett, Gaelyn and Gelbard, Alexander and Ghaffar, Omar and Goldenberg, Anna and Hanna, Karim and Hersh, William and Jain, Jennifer and Jayachandran, Lochana and Jenney, Kaley and Jenkins, Kathy and Jenkins, Phillip and Jo, Stacy and Johnson, Alistair and Kalia, Ayush and Kalia, Megha and Khawa, Zoha and Kobayashi, Kenji and Kostelnik, Cindy and Krause, Alisa and Krussel, Andrea and Lapadula, Elisa and Leo, Genelle and Levinsky, Justin and Loewith, Chloe and Mah, Linda and Maharaj, Vrishni and Maramraju, Sudhiksha and Miao, Siyu and Michaels, LeAnn and Mifsud, Matthew and Mikhael, Marian and Moothedan, Elijah and Nafii, Yosef and Neal, Tempestt and Newberry, Karlee and Ng, Evan and Nickel, Christopher and Peltier, Amanda and Pharr, Trevor and Pontell, Matthew and Potter, Jaiden and Persaud, Jacqueline and Premi-Bortolotto, Claire and Rafatjou, Parnaz and Rahman, JM and Rajkumar, Gayathiri and Rajkumari, Starry and Ramos, John and {de Riesthal}, Michael and Rohde, Sarah and Rossi, Jillian and Russell, Laurie and {Salvi Cruz}, Samantha and Samuel, Joyce and Shah, Suketu and Shawkat, Ahmed and Silberholz, Elizabeth and Stark, John and Su, Lala and Sudhakar, Shrramana Ganesh and Sutherland, Duncan and {Swarna Mukhi}, Venkata and Tang, Jeffrey and Taylor, Luka and Toghranegar, Jamie and Tu, Julie and Urbano, Megan and Vinson, Kimberly and Wilke, Jordan and Wilson, Claire and Zanin, Madeleine and Zeng, Xijie and Zesiewicz, Theresa and Zhao, Robin and Zisimopoulos, Pantelis and Ghosh, Satrajit},
  title = {{Bridge2AI-Voice Pediatric Dataset}},
  journal = {{PhysioNet}},
  year = {2026},
  month = oct,
  note = {Version 2.0.0},
  doi = {10.13026/3ymv-t964},
  url = {https://doi.org/10.13026/3ymv-t964}
}

@article{pollard_physionet_2026, title = {{PhysioNet} as a global platform for biomedical research}, volume = {1}, issn = {3005-0693}, url = {https://doi.org/10.1038/s44360-026-00096-z}, doi = {10.1038/s44360-026-00096-z}, abstract = {Twenty-five years after its first public repository was established, the PhysioNet resource for biomedical and clinical data has enabled tens of thousands of research papers. Further efforts are ongoing to advance secure data sharing, community engagement and scalable research infrastructure.}, number = {8}, journal = {Nature Health}, author = {Pollard, Tom and Moody, Benjamin E. and Lehman, Li-wei H. and Gow, Brian J. and Fernandes, Chrystinne and Xie, Chen and Johnson, Alistair and Mark, Roger G. and Heldt, Thomas}, month = aug, year = {2026}, pages = {792--795},}

Abstract

The human voice contains complex acoustic markers which have been linked to important health conditions including dementia, mood disorders, and cancer. When viewed as a biomarker, voice is a promising characteristic to measure as it is simple to collect, cost-effective, and has broad clinical utility. Recent advances in artificial intelligence have provided techniques to extract previously unknown prognostically useful information from dense data elements such as images. The Bridge2AI-Voice project seeks to create an ethically sourced flagship dataset to enable future research in artificial intelligence and support critical insights into the use of voice as a biomarker of health. Here we present Bridge2AI-Voice, a comprehensive collection of data derived from voice recordings with corresponding clinical information.

Bridge2AI-Voice Pediatric Dataset v2.0.0 contains derived audio features for 41,831 recordings (59.5 hours of audio) collected from 656 participants aged 2–17 at a single site. The release contains data considered low risk, including derivations such as spectrograms but not the original voice recordings. Detailed demographic, clinical, and validated questionnaire data are also made available.

Background

Patients/healthy volunteers at the Hospital for Sick Children were considered for enrollment in the study. Patients were considered eligible for the study if they fulfilled the inclusion criteria of 2 to 18 years of age, and English proficiency. Exclusion criteria included participants over 18 years of age, and individuals who were non-verbal. Non-patients, recruited through research postings, were evaluated for eligibility based on the study’s inclusion and exclusion criteria. Following confirmation of eligibility, parental or participant consent was obtained prior to data collection and data sharing. Once consented, patients were assigned a unique study identification number and a standardized age-appropriate protocol for data collection was adopted. The protocol included the collection of demographic information, voice and speech related questionnaires, and questionnaires inquiring about medical history.

All data was collected through customized software – reproschema-ui – on tablets. A headset was used to record for most participants, while the remaining recordings utilized the built-in tablet microphone due to low tolerance of wearing headphones or existing complex medical conditions. All participants completed the recording and demographic data collection in one session. For participants without adequate comprehension and familiarity with their past medical history, parents or decision-makers completed the survey during the recording on a separate tablet. The simultaneous completion of the recording and survey improved efficiency and minimized participant burden and fatigue. Data were exported, organized into a BIDS-like folder structure with tab-delimited phenotype tables, and de-identified and bundled using b2aiprep, an open-source library developed by our team [4]. Acoustic features were computed with senselab [10].

Recordings were manually reviewed for the voices of people other than the participant and for personally identifying information. Recordings flagged in this review, and recordings not yet reviewed, are excluded from this release. Every released free-speech recording was individually listened to before release.

Participants are identified by a six-digit pseudonym that is the same across versions of this dataset. Sessions are numbered per participant. Calendar dates and times were removed; timing is given instead as intervals, such as days since a procedure or seconds since the first session, or as the local hour of day. Free-text responses were removed. Participants whose answers were judged to carry a high risk of re-identification were not released. To make speaker identification more difficult, the time axis of the spectrogram, Mel spectrogram and MFCC features was down-sampled by keeping every second frame.

Methods

Patients/healthy volunteers at the Hospital for Sick Children were considered for enrollment in the study. Patients were considered eligible for the study if they fulfilled the inclusion criteria of 2 to 18 years of age, and English proficiency. Exclusion criteria included participants over 18 years of age, and individuals who were non-verbal. Non-patients, recruited through research postings, were evaluated for eligibility based on the study’s inclusion and exclusion criteria. Following confirmation of eligibility, parental or participant consent was obtained prior to data collection and data sharing. Once consented, patients were assigned a unique study identification number and a standardized age-appropriate protocol for data collection was adopted. The protocol included the collection of demographic information, voice and speech related questionnaires, and questionnaires inquiring about medical history.

All data was collected through customized software – reproschema-ui – on tablets. A headset was used to record for most participants, while the remaining recordings utilized the built-in tablet microphone due to low tolerance of wearing headphones or existing complex medical conditions. All participants completed the recording and demographic data collection in one session. For participants without adequate comprehension and familiarity with their past medical history, parents or decision-makers completed the survey during the recording on a separate tablet. The simultaneous completion of the recording and survey improved efficiency and minimized participant burden and fatigue. Data were exported and converted to tab delimited values using an open source library developed by our team [4].

Data Description

The dataset contains derived audio features (under features) and phenotypic information acquired during data collection (under phenotype).

Features

Binary files are made available as Parquet, an open-source column-oriented data file format. Each of the parquet files is formatted similarly. Each row contains a unique identifier for the participant (participant_id), the session number within that participant (session_id; sessions are numbered 01, 02, … for each participant, so a session is identified by participant_id and session_id together), the task performed (task_name), the number of time frames for that feature (n_frames), and the tensor data for the feature.

• torchaudio_spectrogram.parquet (n=41,831) contains log-magnitude spectrograms (dB) of dimension 201xT generated with the short-time Fourier transform using a 25 ms window, 10 ms hop length and a 400-point FFT. Every second frame was then dropped, giving an effective hop of 20 ms (50 frames per second).
• torchaudio_mel_spectrogram.parquet (n=41,831) contains Mel spectrograms (power) of dimension 60xT generated with a 25 ms window, 10 ms hop length, a 400-point FFT and 60 Mel bins. Every second frame was then dropped, giving an effective hop of 20 ms (50 frames per second).
• torchaudio_mfcc.parquet (n=41,831) contains 60 Mel-frequency cepstral coefficients of dimension 60xT computed from Mel spectrograms with a 25 ms window, 10 ms hop length, a 400-point FFT and 60 Mel bins. Every second frame was then dropped, giving an effective hop of 20 ms (50 frames per second).
• torchaudio_pitch.parquet (n=41,831) contains the detected pitch (fundamental frequency) over time, of dimension T at 100 frames per second, with a minimum and maximum pitch of 80 and 500 Hz.
• sparc_ema.parquet (n=41,828) contains the estimated electromagnetic articulography (EMA) from a deep learning model [12], of dimension Tx12 at 50 frames per second, where the 12 channels are the X/Y positions of six articulators: tongue dorsum (TD), tongue body (TB), tongue tip (TT), lower incisor (LI), upper lip (UL) and lower lip (LL).
• sparc_loudness.parquet (n=41,828) contains the estimated loudness, the average absolute amplitude of the z-scored audio waveform in 20 ms windows, of dimension T at 50 frames per second.
• sparc_periodicity.parquet (n=41,828) contains the estimated periodicity (confidence of pitch presence), of dimension T at 50 frames per second.
• sparc_pitch.parquet (n=41,828) contains the fundamental frequency (F0) estimated with a different algorithm (CREPE [14]), with a range of 50–550 Hz, of dimension T at 50 frames per second.
• ppgs.parquet (n=41,831) contains phonetic posteriorgram [13] probabilities across 40 categories (39 phonemes and silence), of dimension 40xT at 100 frames per second.

SPARC features could not be computed for 3 recordings (two longer than about 290 seconds and one of 0.34 seconds); these recordings are absent from the SPARC files only.

The features folder also contains two tab-separated files. static_features.tsv (n=41,831) contains one row per recording of features derived from openSMILE [5], Praat [6], parselmouth [7] and torchaudio [8, 9], and a transcription column with an automatic transcript of the recording made with Whisper large-v3-turbo [11]. audio_quality_metrics.tsv (n=41,831) contains five audio quality measures per recording (proportion clipped, proportion silent, peak SNR, spectral-gating SNR and amplitude-modulation depth), each with a pass/fail check.

Phenotype

The phenotype subfolder contains organized information collected from the participant or a parent or guardian during their encounter:

.
├── enrollment
│   ├── eligibility.json
│   ├── eligibility.tsv
│   ├── enrollment_form.json
│   ├── enrollment_form.tsv
│   ├── participant.json
│   └── participant.tsv
├── pediatric
│   ├── pediatric_demographics.json
│   ├── pediatric_demographics.tsv
│   ├── pediatric_medical_conditions.json
│   └── pediatric_medical_conditions.tsv
├── pediatric_questionnaire
│   ├── pediatric_hqa.json
│   ├── pediatric_hqa.tsv
│   ├── pediatric_vhi10.json
│   ├── pediatric_vhi10.tsv
│   ├── pediatric_voice_outcome_survey.json
│   ├── pediatric_voice_outcome_survey.tsv
│   ├── pediatric_voice_related_qol_survey.json
│   └── pediatric_voice_related_qol_survey.tsv
└── task
    ├── acoustic_task.json
    ├── acoustic_task.tsv
    ├── recording.json
    ├── recording.tsv
    ├── session.json
    └── session.tsv

The enrollment folder holds participant, eligibility and enrollment-form tables; pediatric holds demographics and medical conditions; pediatric_questionnaire holds the pediatric PHQ-A, VHI-10, Voice Outcome Survey and Voice-Related Quality of Life Survey; and task holds sessions, acoustic tasks and recordings. The recording.tsv file holds per-recording metadata, including the task, instructions, stimulus and audio format.

enrollment/participant.tsv lists every participant in this release, one row each. While we try to have complete data for every participant, this is not always possible: a consented participant is withheld only when we have no phenotypic data or no audio for them. The other tables have a row for each participant and session in which that form was filled in, so some have fewer rows than participant.tsv: 3 participants have no data on the medical conditions form (pediatric/pediatric_medical_conditions.tsv, 653 rows) and 1 has none on the VHI-10 (pediatric_questionnaire/pediatric_vhi10.tsv, 655 rows).

enrollment/eligibility.tsv and enrollment/enrollment_form.tsv come from the study's eligibility and enrollment forms. In this version they carry little information. No participant is a control (is_control_participant is No for everyone). The only filled enrollment columns, ef_primary_language and ef_select_language, are English for everyone, as is selected_language in participant.tsv. Both tables are kept because participants collected with the study app, for whom these forms are filled, will be added in later versions.

The enrollment, eligibility, session and recording forms are shared with the adult study, so these tables have the same columns in both datasets. Some columns have no values in this version:

• Recorded only by the study app. All data in this version were collected before the app was used, so these columns are empty:
  • enrollment/eligibility.tsv: self_reported_autism, self_reported_speech_delay. Autism is recorded in peds_mc_psych_disorders___asd, and voice disorder or speech delay in peds_mc_v_dis, both in pediatric/pediatric_medical_conditions.tsv.
  • enrollment/enrollment_form.tsv: ef_duration, ef_any_cognitive_challeges, ef_any_hearing_related_req, ef_any_physical_challenges, ef_any_vision_related_req, and ef_fluent_languages___1 to ef_fluent_languages___5.
  • task/session.tsv: session_assigned_tasks.
  • task/recording.tsv: recording_input_gain, recording_microphone, recording_profile_name, recording_profile_version, recording_device_info, recording_device_type, recording_platform, recording_platform_version, recording_screen_resolution, recording_window_resolution.
• Not collected for pediatric participants, and kept for compatibility with the adult data: recording_browser in task/recording.tsv, recorded in the adult data for recordings made in a web browser.
• stimulus_asset_n_images in task/recording.tsv is filled only for stimuli made of several images, where stimulus_asset holds a URL pattern with {n}. Every pediatric image stimulus is a single image, so this column is empty and each stimulus_asset URL opens its image directly.

data_explorer.html, at the top level of the dataset, summarizes every phenotype field: its description, how often it is blank, and the distribution of its values. It opens in any web browser.

Data dictionaries

Every data file has a data dictionary with the same file stem and a JSON suffix (e.g. torchaudio_spectrogram.json). For static_features and audio_quality_metrics, keys are column names and values give a one-sentence description. For the parquet files, the dictionary lists each field with its type and description, and gives the tensor's shape, units, frame rate, extraction parameters and software version. For phenotype tables, the dictionary gives the source form and, for each column, its description, value type and response options.

The code used to preprocess the raw audio waveforms into the feature files and to merge the source data into the phenotype files has been made open source in the b2aiprep library [4].

Usage Notes

If using Python, the parquet dataset can be loaded in with the HuggingFace datasets library as follows:

from datasets import Dataset
ds = Dataset.from_parquet("features/torchaudio_spectrogram.parquet")

Loading a file this way writes a cached copy of it, about the size of the file, to the HuggingFace cache directory.

The Mel spectrogram is stored as power and can be plotted in decibels after conversion. The spectrogram in torchaudio_spectrogram.parquet is already in decibels and can be plotted without conversion.

from datasets import Dataset
import numpy as np
import librosa
import matplotlib.pyplot as plt

ds = Dataset.from_parquet("features/torchaudio_mel_spectrogram.parquet")
mel_db = librosa.power_to_db(np.asarray(ds[0]["mel_spectrogram"]))
plt.figure(figsize=(10, 4))
plt.imshow(mel_db, aspect="auto", origin="lower")
plt.title("Mel spectrogram")
plt.xlabel("Frame (20 ms)")
plt.ylabel("Mel bin")
plt.colorbar(label="dB")
plt.show()

A phenotype file can be loaded with any statistical analysis tool. Participant and session identifiers contain leading zeros, so read them as text; otherwise they are converted to numbers and no longer match the identifiers in the parquet files. In the phenotype tables the session column is named session_id or ends in _session_id (for example peds_demographics_session_id or recording_session_id). For example, with the pandas library in Python:

import pandas as pd

demographics = pd.read_csv(
    "phenotype/pediatric/pediatric_demographics.tsv",
    sep="\t",
    dtype={"participant_id": str, "peds_demographics_session_id": str},
)

Each feature row matches one row of recording.tsv on participant, session and task, so per-recording metadata can be joined to the features:

static = pd.read_csv(
    "features/static_features.tsv", sep="\t", dtype={"participant_id": str, "session_id": str}
)
recordings = pd.read_csv(
    "phenotype/task/recording.tsv", sep="\t", dtype={"participant_id": str, "recording_session_id": str}
)
merged = static.merge(
    recordings,
    left_on=["participant_id", "session_id", "task_name"],
    right_on=["participant_id", "recording_session_id", "task_name"],
)

Release Notes

B2AI-Voice Pediatric v2.0: 357 new participants have been added and one v1.1 participant was withheld, for 656 in total. Features for all free speech tasks were checked for presence of unconsented speakers and PII. This version also changes identifiers, task names and the file layout, so code written for v1.1 needs updating:

• Data source: the pediatric data have been imported into the study's REDCap project, and v2.0 is built from a REDCap export rather than directly from the reproschema-ui data used for v1.1. This accounts for many of the changes below. All released participants were collected with reproschema-ui, not with the Bridge2AI-Voice app; the REDCap-based format allows pediatric data collected with the app to be included in a future version.
• Participants: 299 of the 300 v1.1 participants are included with the same participant_id. One v1.1 participant was withheld because their answers were judged to carry a high risk of re-identification.
• Recordings: 41,831 (v1.1: 23,533). 198 recordings released in v1.1 are not linked to a record in the REDCap project and are not included in this version. A further 78 v1.1 recordings belonged to the withheld participant.
• Task names: in v1.1, the number in a task name such as picture-12 or passage-3 counted recordings in the order they were made. It did not identify the same item for every participant, and extra takes shifted the numbers that followed. In v2.0 each task name identifies one item, for example identifying-pictures-12, reading-passage-3 or repeating-words-pond. Task families were also renamed (for example picture → identifying-pictures, sentence → repeating-sentences, favorite-food → conversation-6-plus-favorite-food). recording_id is unchanged, so v1.1 task names and sessions can be mapped to v2.0 by joining the two versions' phenotype/task/recording.tsv files on recording_id.
• Sessions: session_id is now the session number within the participant (01) instead of an 8-character identifier. session.tsv adds session_index, session_local_hour and session_seconds_since_first_session, and no longer has session_is_control_participant. Control status is now given by is_control_participant in the new enrollment/eligibility.tsv; no participant in this dataset is a control.
• Recording metadata: the metadata folder (metadata.parquet) was removed. Per-recording metadata is now in phenotype/task/recording.tsv, which grows from 11 to 31 columns and adds the task name, language, instructions, stimulus, speech type, recording order and local hour, and audio sample rate and channel count. Some of its columns have no values in this version; the Phenotype section lists them with the reason.
• Enrollment: a new phenotype/enrollment folder holds participant.tsv (site, whether any data was collected remotely, selected language), eligibility.tsv and enrollment_form.tsv.
• Demographics: race is now one column per option (peds_race___1 to peds_race___8) instead of a single column. State or province and form duration were added.
• Medical conditions: multi-select answers are now one column per option, and the number of days from each surgery or procedure to the participant's first session was added.
• Features: all features were recomputed for this release, including for recordings released in v1.1, so values for the same recording can differ slightly from v1.1. audio_quality_metrics.tsv now covers every recording (v1.1: 22,080 of 23,533).
• Data dictionaries: feature dictionaries now give units and frame rates, and the PPG dictionary cites the correct reference.

B2AI-Voice Pediatric v1.1: No new participants released in this minor update, but releasing audio features for all free speech tasks that were manually checked for presence of unconsented speakers and PII. Additionally, releases new metrics related to the audio quality of the recordings and per recording metadata information.

B2AI-Voice Pediatric v1.0: This was the first release of the Bridge2AI-Voice Pediatric dataset.

Ethics

Data collection and sharing was approved by the Research Ethics Board at the Hospital for Sick Children.

Acknowledgements

This release would not be possible without the graceful contribution of data from all the participants of the study.

This project was funded by NIH project number 3OT2OD032720-01S1: Bridge2AI: Voice as a Biomarker of Health - Building an ethically sourced, bioaccoustic database to understand disease like never before. We would also like to thank the NIH for their continued support of the project.

Conflicts of Interest

None to declare.

References

1. Kelchner, L. N., Brehm, S. B., de Alarcon, A., & Weinrich, B. (2012). Update on pediatric voice and airway disorders: assessment and care. Current opinion in otolaryngology & head and neck surgery, 20(3), 160–164. https://doi.org/10.1097/MOO.0b013e3283530ecb
2. Tavares, E. L., Labio, R. B., & Martins, R. H. (2010). Normative study of vocal acoustic parameters from children from 4 to 12 years of age without vocal symptoms: a pilot study. Brazilian journal of otorhinolaryngology, 76(4), 485–490. https://doi.org/10.1590/S1808-86942010000400013
3. Fujiki RB, Venkatraman A, Heller Murray ES. The Pediatric Vocal Mechanism: Structure and Function. J Voice. 2025 Apr 4:S0892-1997(25)00118-3. doi: 10.1016/j.jvoice.2025.03.025. Epub ahead of print. PMID: 40187973; PMCID: PMC12353639.
4. Johnson, A., Ng, E., Bevers, I., Wilke, J., Brito, R., Bedrick, S., Catania, F. & Ghosh, S. (2026). Bridge2AI Data Processing Library (Version 4.0.0) [Computer software]. https://github.com/sensein/b2aiprep
5. Florian Eyben, Martin Wöllmer, Björn Schuller: "openSMILE - The Munich Versatile and Fast Open-Source Audio Feature Extractor", Proc. ACM Multimedia (MM), ACM, Florence, Italy, ISBN 978-1-60558-933-6, pp. 1459-1462, 25.-29.10.2010.
6. Boersma P, Van Heuven V. Speak and unSpeak with PRAAT. Glot International. 2001 Nov;5(9/10):341-7.
7. Jadoul Y, Thompson B, De Boer B. Introducing parselmouth: A python interface to praat. Journal of Phonetics. 2018 Nov 1;71:1-5.
8. Hwang, J., Hira, M., Chen, C., Zhang, X., Ni, Z., Sun, G., Ma, P., Huang, R., Pratap, V., Zhang, Y., Kumar, A., Yu, C.-Y., Zhu, C., Liu, C., Kahn, J., Ravanelli, M., Sun, P., Watanabe, S., Shi, Y., Tao, T., Scheibler, R., Cornell, S., Kim, S., & Petridis, S. (2023). TorchAudio 2.1: Advancing speech recognition, self-supervised learning, and audio processing components for PyTorch. arXiv preprint arXiv:2310.17864
9. Yang, Y.-Y., Hira, M., Ni, Z., Chourdia, A., Astafurov, A., Chen, C., Yeh, C.-F., Puhrsch, C., Pollack, D., Genzel, D., Greenberg, D., Yang, E. Z., Lian, J., Mahadeokar, J., Hwang, J., Chen, J., Goldsborough, P., Roy, P., Narenthiran, S., Watanabe, S., Chintala, S., Quenneville-Bélair, V, & Shi, Y. (2021). TorchAudio: Building Blocks for Audio and Speech Processing. arXiv preprint arXiv:2110.15018.
10. Catania, F., Wilke, J., & Ghosh, S. (2026). senselab: a Python package for speech and voice analysis (Version 1.4.0-alpha.0) [Computer software]. https://github.com/sensein/senselab
11. Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2023). Robust Speech Recognition via Large-Scale Weak Supervision. Proceedings of the 40th International Conference on Machine Learning, PMLR 202:28492–28518. arXiv:2212.04356
12. Cho, C. J., Wu, P., Prabhune, T. S., Agarwal, D., & Anumanchipalli, G. K. (2024). Coding Speech through Vocal Tract Kinematics. IEEE Journal of Selected Topics in Signal Processing, 18(8), 1427. https://doi.org/10.48550/arXiv.2406.12998
13. Churchwell, C., Morrison, M., & Pardo, B. (2024). High-Fidelity Neural Phonetic Posteriorgrams. ICASSP 2024 Workshop on Explainable Machine Learning for Speech and Audio. https://doi.org/10.48550/arXiv.2402.17735
14. Kim, J. W., Salamon, J., Li, P., & Bello, J. P. (2018). CREPE: A Convolutional Representation for Pitch Estimation. 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 161–165. arXiv:1802.06182