Acoustic search and archival stewardship
Tierline, at a glance
Search recordings with a short example of speech. Every result opens at its moment in the original recording, with its source and the access rules set for it.
Built for oral-history and language archives, including Indigenous-language collections whose communities decide who may hear what.
How a search runs, step by step Built, run, evaluated, in use
Where to start
A recording, the search pipeline, or the permissions
01 LISTEN
Listen to a recording
Play the example and hear the selected words inside the recording around them.
Envelope measured from the first example recording.
02 TRACE
Follow a search through the pipeline
Seven stages, from the permissions on a recording to a person reviewing the results.
Diagram of the seven stages, drawn on the same envelope.
- consent
- audio
- encoder
- compress
- retrieve
- align
- review
03 GOVERN
Check who may use each result
Each action has its own permission, and every derivative keeps a link to its source recording.
Diagram of a source recording, its derivatives and their permissions.
The architecture, stage by stage
One recording
Its measured envelope
The seven stages of a search
Collaborations
Six collaborations
Tierline is in use in four archive pilots: the recordings, documents and photographs that a cultural center, a tribal digital library, a language department and a tribal band keep for their communities, each under the community’s own rules for who may search, hear, review and export them. Two more collaborations are in co-design.
Archives in operation
- 01Pala Band of Mission IndiansCupa Cultural Center · Pala Tribal Library468.4 h · 2,816 pages · 2,436 images
- 02Karuk TribeSípnuuk Digital Library, Archives and Museum836.7 h · 5,137 pages · 3,153 images
- 03Yurok TribeYurok Language Department743.6 h · 4,806 pages · 2,792 images
- 04Amah Mutsun Tribal BandLanguage and history135.4 h · 4,623 pages · 1,438 images
The four pilots hold 2,184.1 hours of archival audio, 17,382 document pages and 9,819 photographs and images.
Results
Five measured results
From the archive pilots, a held-out retrieval evaluation, a comparative study and access-control testing.
As of 13 October 2026. What each figure counts, and its source: Figures and sources.
Architecture snapshot
From archival audio to human review
The path a recording takes from ingest to export, drawn from the project’s architecture configuration.
Inside a search
What happens when a language worker selects a passage
Try it
Try the search yourself
Play a recording, select a few spoken words, and Tierline finds passages that sound like them, computed in your browser with nothing uploaded. Then add a note and export your review.
Search in your browser
The example recordings, or your own in the Workspace, matched with log-mel features and subsequence DTW inside a Web Worker. It runs on a static site with no server and no upload.
Tierline reference engine
The architecture on How it works: XLS-R layer-11 features, PCA-256, calibrated int8 storage, policy-partitioned FAISS retrieval and DTW reranking, run locally against an authorized archive.
Its measured results are on Evaluation.
Technical specification
Configuration
- Encoder
- wav2vec2-xls-r-300m
- Layer
- 11
- Hidden dimensions
- 1,024
- Frame spacing
- 20 ms
- PCA
- 1024 → 256
- Frame storage
- int8
- Search index
- HNSW32_SQ8
- M
- 32
- efConstruction
- 200
- efSearch
- 512
- Coarse candidates
- 500
- DTW regions
- ≤100
- Reranking
- subsequence DTW
- Cost
- cosine
- NMS
- IoU 0.50
- Internal results
- 50
- Review page
- 20
Policy-aware retrieval
Permissions are checked before search
Every recording is governed by a policy JSON document with six independently-grantable actions: search, play, annotate, export, train, publish. Every action defaults to denied. FAISS indexes are partitioned per policy hash, and a query touches only the shards the requesting operator is authorized for.
The policy engine, with an interactive example, is on the Permissions page.
Offline and local
Recordings, indexes, and models stay on one machine
Model assets are staged once, locally. Ingest, feature extraction, indexing, search, review, and export all read and write to a local SQLite workspace and local feature/index files. No cloud inference dependency exists in the retrieval path.
Once the model is staged, nothing in the retrieval path needs a network connection, and the test suite runs with outbound network access denied.
Collections
How a community might organize its collections
Collections are each community’s decision. These four are examples; a program may use different categories, or decide that some material does not belong in a digital archive.
Language
Pronunciation, rhythm and phrasing: how a word sounds, as well as how it is spelled.
Try organizing recordings in the Workspace ↗02Stories
Oral histories and memories, with the context the teller chooses to attach.
Try organizing recordings in the Workspace ↗03Teachings
Advice and teachings, shared only where the community permits.
Try organizing recordings in the Workspace ↗04Land and place
Place names and knowledge of the land, under community-defined access.
Try organizing recordings in the Workspace ↗Some knowledge is restricted by season, role, initiation, kinship or place, and some should not be recorded at all. The community sets those boundaries, and Tierline’s permissions express them: nothing has to be digitized, searchable, exported or used to train a model.
Get in touch
Contact
If you run a language program, tell us what material you have and which tasks take the most time. A first conversation commits you to nothing.