stcoats/latrobe-ca-asr-exact-v2

La Trobe CA-ASR Exact-Boundary Dataset

This dataset contains the processed La Trobe Corpus of Spoken Australian English examples used for the principal ASR adaptation experiment accompanying Keeping the Ums: Conversation-Analysis-Informed Verbatim ASR for Web Speech.

Source and licence

The source is the La Trobe Corpus of Spoken Australian English (Mullan, 2002), La Trobe University, <https://doi.org/10.26181/23089559>. The source collection is distributed under the Creative Commons Attribution-NonCommercial 4.0 International licence (CC BY-NC 4.0). This processed derivative is distributed under the same licence: reuse and redistribution must be non-commercial, with attribution to the original corpus and an indication of the changes made.

Processing

The original conversational transcripts were normalized into a consistent lexical-verbatim convention. The target text always comes from these normalized human transcripts; ASR output was used only to recover approximate word times. Recordings were converted to mono 16-kHz PCM WAV and divided into fixed training, validation, and held-out test partitions. Examples with transcript-marked temporal overlap were excluded from this primary release. The dataset retains fillers, repetitions, cut-offs, reduced forms, and selected audible non-speech events.

ASR-assisted segmentation and exact cuts

Each complete source recording was first transcribed with Faster-Whisper Large-v3 in English (beam size 5, condition_on_previous_text=False, no VAD), after which WhisperX's default English alignment model supplied word-level timestamps. The normalized human transcript and timestamped ASR words were then compared as token sequences. Exact matches belonging to runs of at least two consecutive tokens served as timing anchors. Times for unmatched human transcript tokens were estimated by monotonic linear interpolation between anchors; unmatched token durations used the median ASR-word duration.

Candidate examples targeted approximately 20 seconds. Human-transcript turn boundaries were preferred where they produced examples between 10 and 30 seconds; a long individual turn was instead divided at the token closest to the 20-second target. An earlier preparation stage added margins for checking and possible forced alignment. Those margins were removed for this release: each WAV begins at the estimated start of its first target token and ends at the estimated end of its last target token, with no safety padding. Thus, source_start and source_end are WhisperX-derived boundaries: directly aligned where the boundary token was an anchor and interpolated otherwise.

Quality tiers

quality_tier describes confidence in the automatic mapping between the human transcript and the ASR-derived timeline; it is not a subjective rating of recording quality. anchor_coverage is the proportion of human-reference tokens belonging to exact-match anchor runs. asr_similarity is the SequenceMatcher similarity between the human-reference span and the nearby ASR words.

• —A required anchor coverage of at least .65, ASR similarity of at least .70, no transcript-marked overlap or uncertain material, and no multiple-speaker flag.
• —B required anchor coverage of at least .50, ASR similarity of at least .55, transcript-marked overlap of no more than 12% of tokens, and no uncertain material.

The final primary release then excluded every example containing any transcript-marked temporal overlap, independently of tier. It contains 208 tier-A and 30 tier-B examples. Tier B therefore mostly identifies lower anchor/similarity confidence or sequential speaker changes in this release, not retained simultaneous speech.

The release contains 238 examples (1.189 hours): 136 training examples, 25 validation examples, and 77 held-out test examples. The test partition was not used for model selection. Splits are recording-disjoint: HeatherMarie is the validation recording, NatalieKen is the test recording, and the remaining four recordings constitute training data.

Fields

Each split uses the Hugging Face AudioFolder layout with a metadata.csv file. Fields include the WAV filename, normalized verbatim text, example and source recording identifiers, speaker labels, duration, source-recording offsets, token count, quality tier, and boundary method.

Attribution

Users must cite the original La Trobe corpus and indicate that this release is a segmented and normalized derivative. For the preparation method and accompanying ASR adaptation experiment, please also cite:

Coats, Steven. (2026, forthcoming). Keeping the Ums: Conversation-Analysis-Informed Verbatim ASR for Web Speech. Proceedings of the 13th Web-as-Corpus Workshop (WAC 2026). Accepted for publication.

Limitations

The release is small and was prepared for verbatim ASR adaptation rather than as a representative sample of Australian English. Speaker changes can occur within an example; the exclusion rule concerns transcript-marked simultaneous overlap. Transcript conventions have been normalized, so consult the original corpus for the complete conversation-analytic annotation.