CoolFace
Datasetpublic

ananth-r-gnani/aura-phone-dictation-eval

Aura Phone Dictation Eval Evaluation set of 365 progressive audio clips from 142 phone-number dictation sequences extracted from Aura Hindi/English call-center recordings. This dataset is used to evaluate end-of-turn (EOT) detection models on structured phone-number dictation. Each sequence captures a caller dictating a 10-digit Indian mobile number across multiple speech segments. Progressive clips accumulate earlier segments plus trailing silence, ending with a final clip once… See the full description on the dataset page: https://huggingface.co/datasets/ananth-r-gnani/aura-phone-dictation-eval.

sourceHugging Faceotherupdated 18d agoView on Hugging Face
0likes50downloads
Dataset Card

Aura Phone Dictation Eval

Evaluation set of 365 progressive audio clips from 142 phone-number dictation sequences extracted from Aura Hindi/English call-center recordings.

This dataset is used to evaluate end-of-turn (EOT) detection models on structured phone-number dictation. Each sequence captures a caller dictating a 10-digit Indian mobile number across multiple speech segments. Progressive clips accumulate earlier segments plus trailing silence, ending with a final clip once all 10 digits have been spoken.

What this dataset contains

  • —Source: Aura hi-IN call recordings (dictation_clips_stereo/)
  • —Task: Detect whether a progressive dictation clip is EOT (complete 10-digit number spoken) or NOT_EOT (dictation still in progress)
  • —Labels: GPT-OSS progressive EOT labels after clean 10-digit curation
  • —Clips: 365 total — 144 EOT, 221 NOT_EOT
  • —Sequences: 142 verified 10-digit phone dictation sequences
  • —Audio: Mono WAV clips, max 15 s, extracted with Silero VAD segmentation

Extraction pipeline

  1. 1.Scan Aura calls for structured phone-number dictation patterns
  2. 2.Extract stereo dictation sequences with progressive NEOT/EOT WAV clips
  3. 3.Verify sequences with Vachana STT + GPT-OSS quality checks
  4. 4.Curate clean sequences where digits progress monotonically to exactly one 10-digit mobile number (no full-number repetition or ambiguous hesitations)
  5. 5.Label each progressive step EOT/NOT_EOT based on cumulative transcript

Fields

ColumnDescription
audioProgressive dictation clip (WAV)
clip_idUnique clip key `{sequence_id}\{clip_name}`
sequence_idDictation sequence identifier
call_idSource Aura call UUID
group_nameAura corpus group (e.g. Choice, IIFL)
stepProgressive step index (1-based)
n_steps_totalTotal progressive steps in sequence
eot_labelGround truth: EOT or NOT_EOT
progressive_vachanaCumulative Vachana STT transcript up to this step
progressive_corpusCumulative corpus ASR transcript up to this step
reconstructed_10digitTarget 10-digit mobile number for the sequence
is_finalWhether this is the final progressive clip
duration_sClip duration in seconds
label_reasoningGPT-OSS reasoning for the EOT label

Optional model score columns (st_v32_prob, phone_ft_prob, etc.) are included when available from local eval runs.

Usage

python
from datasets import load_dataset, Audio

ds = load_dataset("ananth-r-gnani/aura-phone-dictation-eval", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=16_000))

example = ds[0]
print(example["eot_label"], example["progressive_vachana"], example["reconstructed_10digit"])

Intended use

  • —Evaluation only — benchmark EOT / smart-turn models on phone-number dictation
  • —Not intended for training without additional curation and consent review

Citation

Extracted and curated by Gnani.ai from Aura call recordings for internal EOT model evaluation.