The Nixon Corpus: Speaker-Annotated Transcripts of the Nixon Tapes Using a Fine-Tuned Transformer Model

Between February 16, 1971 and July 18, 1973, the Nixon White House secretly recorded 2,719 hours of conversations in the Oval Office, the Executive Office Building, the Cabinet Room and Camp David, and on presidential phone calls. Hand transcription reached only the conversations needed for the Watergate trials, diplomatic histories and edited selections. The National Archives (NARA) has digitized and released all the recordings, but an annotated, accurate transcript of all the audio files has not yet been achieved. Here we present the Nixon Corpus: all 19,426 individual recordings transcribed, 2,206 hours, 19.0 million words, every word timed and every speech segment labeled. To achieve this, a fine-tuned speech recognizer model (Parakeet) was combined in tandem with an independent model (Whisper) to generate two transcripts per conversation, while another voiceprint classifier trained on the tapes assigned speakers to each word. A large language model (Gemini 4 Argon) harmonized these data by taking into context the National Archives' subject logs from each respective tape and Chief of Staff H. R. Haldeman's diaries from that day and the following day. On 165 held-out conversations transcribed by the State Department's historians, the corpus recovered 76.1% of words and 80.1% of names, against 69.2% and 60.0% for Whisper alone. The Nixon Corpus can be used to trace how certain policies developed in the Nixon administration. We have identified several potentially novel histories, such as asking for campaign donations for an ambassadorship and conversations surrounding acting FBI Director L. Patrick Gray. Corpus (release 1.2): https://doi.org/10.5281/zenodo.23657174. Code and paper source: https://github.com/Akoopie/nixon-corpus. Corpus viewer: https://nixoncorpus.org