Research & team

Three research contributions, and what each has achieved

Tierline’s research is about searching speech that belongs to someone: keeping the search fast without letting it reach what a person may not hear, comparing passages rather than transcripts, and making a withdrawal hold even against old indexes.

Permission-aware approximate retrieval for governed archival collections

The approximate index is partitioned by access policy, so a search only ever opens the partitions the person is authorized to search; candidates from material they may not reach are never generated.

99.04% candidate-set Recall@500 against exact search
0 unauthorized disclosures in 12,486 adversarial access cases
Recall@500 is from the approximate-versus-exact index benchmark; the access figures are from automated tests in a test environment.

In the code: src/tierline/retrieval/faiss_index.py, src/tierline/policies/evaluator.py.

Multi-example spoken-passage matching and controlled reranking

Candidate regions proposed by the index are re-scored frame by frame with subsequence dynamic time warping, within bounded region counts and duration ratios, so the final ranking compares the passages themselves. Up to five human-confirmed examples of a passage can be searched together, each on the same bounded path, keeping the closest overlapping match.

10.77 points Hit@5 above frozen XLS-R embeddings
72.76% of positive queries answered by the first result
Both figures are from the held-out evaluation, whose configuration includes bounded reranking; search with several confirmed examples uses the same reranking path.

In the code: src/tierline/retrieval/multi_example.py, src/tierline/retrieval/dtw.py.

Provenance-linked retrieval with tested withdrawal and revocation

Every result is bound to its source recording, derivative and policy; a withdrawal or revocation is enforced at retrieval time, including against stale indexes and cached results.

100% of tested revoked-access cases denied retrieval
27 revocation and withdrawal exercises, including stale indexes and caches
From controlled exercises in a test environment.

In the code: src/tierline/ingest/provenance.py, src/tierline/stewardship.

The full evaluation → How the system works →

Next studies

The questions the next studies take up.

How retrieval accuracy varies from one language and collection to another, above all for languages absent from XLS-R’s pretraining.
Whether contrasts such as ejectives and glottalized sonorants survive PCA-256 projection and int8 quantization.
Whether articulatory-feature-weighted edit distance outperforms the uniform baseline for phone-level retrieval; both are implemented.
How much community-supplied adaptation improves retrieval, and in what form that time is best spent.
Which tasks in a language program’s daily work search by example helps with most, beyond finding a known passage, which is what the comparative study measured.

The team

Twenty-one people, most of them from the communities.

A student founder leads the engineering, with three more engineers and two research collaborators. Twelve Indigenous community stewards and reviewers shape collections, permissions, metadata and evaluation, and all twelve are compensated or supported by their institutions. Two preservation contractors and a technical evaluator, who reviews the quantitative results, complete it.

The 21 contributors and professional collaborators, one mark each.

community-side deployment leads

academic methods consultations

community coordinators

People are named only with their approval.

Write with a research or technical question →