by Paula Orozco-Espinel, PhD Candidate in History
As established in the previous entry, oral history involves a lot of slow work. From recording to archiving and disseminating interviews, each step requires significant time and effort. The capacity of AI to assist researchers varies greatly depending on the specific task at hand. While I previously argued against the automation of oral histories’ collection, I have found AI incredibly useful in the processing phase. When it comes to transcribing, indexing, and interpreting oral histories created through traditional methods, AI offers possibilities that can shorten work hours while simultaneously enhancing historical research.
Processing Oral Histories
AI has proven particularly helpful for transcribing oral histories, often the first step in analyzing them. Automatic speech recognition (ASR) software, such as otter.ai, can create transcripts and even summaries in the blink of an eye. However, ASR is far from perfect. In my experience, a human is still necessary for editing the AI-generated version to produce a good transcription. Fortunately, the best ASRs require little enough editing to still be a significant time-saver.
Even though some ASRs, including Whisper (an OpenAI product), claim to enable transcription in multiple languages and accents, they carry the biases of the data on which they were trained. Therefore, they do not work equally well with all recordings. When working with narrators from underrepresented populations, accuracy decreases. Unsurprisingly, understanding the Spanish of elderly rural Colombian women is not ASR tools’ strongest point. Whisper can also be used to translate from other languages to English, with quite good results—at least in my experience, using Spanish originals.
Privacy concerns are important to keep in mind when transcribing and translating with the support of software such as Otter and Whisper. The best language models do not process on your computer but operate in the cloud. This means you cannot guarantee narrators that interviews “will never be out of your sight” or that “only you will have access to recordings.” This may not be a big issue sometimes. However, in some cases, privacy concerns are enough to justify manual transcription. My current research project, for instance, also includes interviews with people who offered abortion care when doing so was illegal in Colombia. Even though penal prosecution is no longer a concern given the current legislation, some narrators are still worried about the possible negative consequences of their names and strategies being revealed.
AI can also facilitate systematizing oral history metadata, which can be a time-consuming and challenging but crucial task for tackling specific research questions. For instance, Chris Pandza used AI to make the content of the massive Ellis Island Oral History Project records easier to navigate and interpret. Although information regarding narrators’ birthdays, age of immigration, and country of origin had been systematically recorded in transcript headers, the inability to filter interviews based on these attributes without opening each file posed a challenge to, let’s say, studying Polish immigrants. Moreover, the inconsistent format of the headers made systematizing through pattern recognition impossible (e.g., the spelling of the word “birthday” and the format of dates varied). As a result, Pandza manually recorded metadata from 100 documents, fed these documents and his results into GPT-3, and then used the trained GPT-3 model to process the entire corpus of 2,000 interview transcripts. This saved him over 150 hours of otherwise tedious work!
Nowadays, it is not even necessary to know how to code to do something similar to what Pandza did, as he notes himself. The OpenAI product Playground enables the use of plain language, thus allowing oral historians without training in computer science to let their research interests lead them through fruitful explorations. You can not only organize metadata but also automatically create indexes, identify topics in transcripts, create timestamps, and more. I have not yet systematically used OpenAI Playground to process my oral histories, but I have found it useful to occasionally help me find quotes I could not recall well enough to easily find through keyword search. Keep in mind a single 2-hour interview generally result in a 40-single-space-pages-long transcript.
Thus, while AI may not be the best tool for collecting oral histories, its potential to improve the processing and analysis phases is immense. By thoughtfully integrating AI tools into these stages, oral historians can increase efficiency and their ability to collect information from interviews. It is worth continuing to weigh the benefits of using AI for other stages of an oral history project, including augmenting collected voices through the creation of impactful dissemination products.