The AI Museum
AI Analytics1,072 recordsRecounted 23 September 2026

Speech models from United States of America

Weightsopenclosedunknownfrontier only
Year20262025202420232022202120202019201820172016
2 filters active, showing 26 of 1,072 recordsClear all

26

Models

14 organisations

41%

Open weights

7 open, 10 closed

4

Frontier

By Epoch's flag

2B

Largest traceable

SeamlessM4T

1966–2023

Span

14 with a traceable size

Releases

by year, from 2008
2008200920102011201220132014201520162017201820192020202120222023

Organisation

Country

Domain

Records

26
Speech models from United States of America, newest first. Each row links to the source Epoch AI records for it.
ModelOrganisationReleasedParametersWeights
Mamba-24M (SC09)Carnegie Mellon University (CMU)2023-1223MClosed
SeamlessM4TFacebook2023-122BOpen
WhisperOpenAI2022-092B *Open
data2vec (speech)Meta AI2022-01705M *Open
W2v-BERTGoogle Brain2021-081BClosed
HuBERTFacebook AI Research2021-071B *Open
BigSSLGoogle2021-018B *Closed
Conformer + Wav2vec 2.0 + Noisy StudentGoogle2020-101BClosed
wave2vec 2.0 LARGEFacebook2020-10317MOpen
ConformerGoogle2020-05119MClosed
ContextNetGoogle2020-05113M *Closed
ContextNet + Noisy StudentGoogle2020-01—Closed
Transformer - LibriVox + Decoding/RescoringFacebook2019-11296MOpen
Big-Little Net (speech)IBM2018-073M *Open
LF-MMIJohns Hopkins University2016-0917MClosed
MS-ensemble-speech-recognitionMicrosoft2016-093B *Closed
WaveNetGoogle DeepMind2016-09—Not recorded
TC-DNN-BLSTM-DNNCarnegie Mellon University (CMU)2015-0418MClosed
Multilingual DNNGoogle2013-05207MNot recorded
ReLU-SpeechGoogle2013-05102M *Not recorded
DistBelief SpeechGoogle2012-1247M *Not recorded
NETtalk reimplementationOregon State University1990-0627,480Not recorded
Speaker-independent vowel classificationUniversity of Washington1989-113,040Not recorded
NetTalk (dictionary)Princeton University1987-0618,629Not recorded
NetTalk (transcription)Princeton University1987-0618,629Not recorded
LTE speaker verification systemIBM1966-112,061 *Not recorded

An asterisk marks a parameter figure the source does not consider traceable to the people who built the model. That judgement is the source’s and it has false negatives: Kimi K3 is marked speculative at 2.8T while Moonshot’s own model card states the figure. Why that distinction carries the whole exhibit.

Counts are of models Epoch AI records as notable, not of every model released. The most recent month is always incomplete, because a model enters the dataset when it is reviewed rather than when it ships.

Epoch AI, 'Data on AI Models'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/data/ai-models-documentation' Used under CC BY 4.0.

How this museum is made