The problem
Why archival speech is hard to search
Why the speech inside a catalogued collection can still be out of reach. Described from public sources, not from any one archive.
Scale
How much language material California’s archives hold
UC Berkeley's Survey of California and Other Indian Languages reports more than 150 linear feet of archival material, more than 2,000 catalog records, and primary-source material concerning approximately 130 languages, including at least 50 Indigenous languages of California.[1]
A catalog can tell a researcher that a collection exists. It does not necessarily make the speech inside hours of recordings searchable by word, sound, construction, or example.
Tierline has no access to these collections; they are cited for the scale of the problem. Tierline’s held-out evaluation is described, with its scope, on Evaluation.
The everyday version
What this costs a language worker
Some important California collections include field recordings and elicitation materials created during twentieth-century linguistic documentation; the Survey notes that students in the 1950s and 1960s used elicitation protocols across California, and those materials are now part of its permanent collection.[2]
Consider a teacher who wants examples of a particular word, pronunciation or grammatical form somewhere in a long collection. Without time-aligned transcription or an audio index, locating every relevant passage may require extensive manual listening. How much depends entirely on the collection and the query; there is no universal number. What is constant is that the effort scales with the length of the collection rather than the size of the answer — and the resource being spent is the time of people who can do work almost nobody else can do.
Why not transcribe
Why not transcribe everything?
Transcription of extremely low-resource and archival speech remains difficult and can produce fluent output that is wrong in ways a non-speaker cannot detect. Where a transcript is wrong, everything downstream inherits the error invisibly.
Tierline takes a narrower path: retrieve likely matches, narrow what a person must review, leave interpretation to the person. Prior work on query-by-example spoken-term detection for exactly this kind of endangered-language material motivates the approach.[14]