Research
Talio for researchers
Talio is a speaking-practice platform for university language programs. Every conversation a learner has with the AI tutor is stored as a complete record: both speakers, timed, annotated, and comparable across sessions. This page describes what that record contains, what a study can vary, and how to work with us.
What a conversation records
| Signal | Granularity | Speakers |
|---|---|---|
| Transcript | Per turn, from a post-conversation speech recognizer, with a second recognizer run in tandem | Learner and tutor |
| Word timing | Per word and per token | Learner |
| Pauses | Voice-activity spans within a turn; a silence of a quarter second or longer splits a run | Learner |
| Turn timing | Start and end of every turn, so response latency is measurable | Learner and tutor |
| Pronunciation | Per word, per phoneme, and per syllable with tone in Mandarin, from a separate assessment engine | Learner |
| Lemma, part of speech, morphology | Per token: mood, tense, verb form, case, and aspect where the language marks them | Learner and tutor |
| Feedback | Each feedback item with its error type in the language's taxonomy | Learner |
| Can-do evidence | One verdict per statement per conversation | Learner |
| Audio | Retained per conversation | Learner and tutor |
Most learner corpora are monologic or interview-based and untimed, and the interlocutor's speech is not part of the record. Here it is, and the interlocutor is a variable you can set.
What a study can vary
The settings in force for each conversation are stored with its outcomes: the scenario and the target level. A within-learner design that would take a term to arrange with human interlocutors is a scenario setting here. Learners can also be followed across weeks, because every conversation is tied to the same learner record.
Questions this record supports
- Does a learner reuse the tense or construction the tutor used in the previous turn?
- How many times does a learner hear a word before producing it, and does that differ by word class?
- Where do a learner's pauses fall: inside clauses, or between them?
- How does turn-transition latency change over weeks?
- What share of the tutor's words has the learner ever produced?
None of these is a published Talio finding. They are the kinds of question the record supports.
How a study works
- Studies run under your institution's IRB. Talio provides a description of the data for your protocol.
- Learners consent to research use separately from using Talio in class. Students who do not consent still use the tool, and their data is not part of the study.
- Data for a consented study is shared de-identified under a data-use agreement with your institution. Audio is shared only when the protocol and the consent cover it.
- Data processing agreements are available for institutional partners.
- Talio's measures, and what each one does not measure, are documented on the metrics page.
Where Talio is used
In use in language courses at two R1 universities. LaunchPad 2026 finalist (Center for Language & Technology, University of Hawaiʻi at Mānoa, presented at CALICO 2026). SOC 2 Type I audit complete · Type II in progress.
Talk with us about a study
If you are designing a study, we would like to hear about it.