OcularAIInc/AMERICAN-ENGLISH-TRANSCRIBED-HIFI-FULL-DUPLEX-TWO-SPEAKER-CONVERSATIONAL-DATASET-SAMPLE

AMERICAN ENGLISH TRANSCRIBED HI-FI FULL-DUPLEX TWO-SPEAKER CONVERSATIONAL DATASET — SAMPLE

Overview

This open sample from Ocular AI contains four American English conversations between two people, with a separate audio track for each speaker and verbatim transcripts containing segment- and word-level timestamps.

The recordings capture conversational exchanges: repetitions, fillers, false starts, pauses, laughter, and audible breaths. Some conversations begin with a supplied discussion prompt; this is not a claim that every utterance was unprompted. Separate speaker tracks let researchers examine both sides of an exchange and reconstruct their shared timeline without relying on a single mixed recording.

This release is the four-conversation back-and-forth package supplied for the Ocular–Cekura evaluation collaboration. It is a sample of a larger Ocular collection, not the complete collection and not a standalone reproduction of the benchmark. It contains recordings and reference annotations, not model predictions or benchmark score files. It is also distinct from Ocular's previously shared private one-hour sample, which used a different selection of conversations.

Contents

Statistics below were calculated from the actual files in this release.

Property	Value
Conversations	4
Speaker tracks / metadata rows	8, two per conversation
Distinct speaker IDs	8
Conversation timeline duration	60.41 minutes total
Combined duration of all speaker tracks	120.64 minutes total
Approximate duration per conversation	15 minutes
Transcript segments	300
Word/event entries in segments[].words	12,970; includes non-speech markers, not only lexical words
Language / locale	English / en-US
Audio layout	FLAC, mono track per speaker
Split	train (a packaging convention, not a recommended evaluation split)

Conversation duration is the latest track end after applying relative recording offsets, summed across sessions. Summing both speaker tracks produces roughly twice as many minutes and should not be reported as conversation duration.

Collection and transcription

Participants were recorded on separate microphones during two-person conversations. Speaker demographics and accent labels are self-reported. Topics include interpreting news, drones and their implications, customer engagement, and fitness decisions. Conversation prompts may appear in the transcript alongside the ensuing discussion.

According to the source collection's documentation, transcripts were produced through an ASR-assisted process followed by human transcription correction, review, and Ocular auditing. Word timestamps were generated by forced alignment against reviewed text. The references retain conversational features rather than rewriting speech into polished prose. This does not mean every boundary or token is error-free.

Repository layout

README.md
LICENSE
train/
  metadata.jsonl
  <speaker-track-1>.flac
  ...
  <speaker-track-8>.flac

metadata.jsonl contains one JSON object per speaker track. file_name references the corresponding FLAC file in the same directory. The layout follows Hugging Face AudioFolder conventions so the audio and annotations can be used together. Original source manifests and per-track JSON exports are not included: their retained content has been consolidated into the release schema below.

Metadata schema

There are 14 top-level fields in the raw metadata. The AudioFolder loader may represent file_name as an audio feature rather than retaining the original field verbatim.

Field	Type	Description
session_id	string	Identifies the conversation; exactly two tracks share each session ID.
file_name	string	Relative FLAC filename for this row.
speaker_id	string	Source contributor identifier; eight distinct values in this sample.
take_id	string	Source identifier for this individual recording.
language	string	Language code, en.
language_locale	string	Locale code, en-US.
datetime_utc	string	Source recording timestamp in UTC, preserved from the delivery metadata.
duration_seconds	number	Duration of this speaker track, in seconds.
gender_self_reported	string	Participant's self-reported gender label.
age_range	string or null	Self-reported age band; missing for four of eight tracks.
native_language	string	Self-reported native language.
accent_self_reported	string	Self-reported accent label.
transcript	string	Full speaker transcript, with conversational disfluencies and inline event tags.
segments	list of objects	Timestamped transcript segments and their word/event entries.

Each segment has index (integer), start and end (seconds), text (string), and words (list). Each word/event entry has word (string), start, and end (seconds). Times are relative to that speaker's own audio file, not to its partner's file and not to UTC.

The public metadata intentionally omits vendor_id, voice_self_description, the complete vendor_qc object (including passed, issues_flagged, and notes), sample_rate_hz, and bit_depth. Removing those metadata columns does not resample or alter the audio: codec properties remain readable from the files. Original source IDs, timestamps, filenames, and audio synchronization tags are retained; this is not an anonymized recording release.

Verbatim conventions and non-speech events

The references preserve repetitions, hesitations, restarts, and event annotations. Tags should not be treated as ordinary spoken words without an explicit task-specific policy.

Tag	Meaning / usage
<breath/>	Audible breath; 616 occurrences in the full transcripts.
<laugh/>	Standalone laughter; 108 occurrences.
<laugh> … </laugh>	Paired laughter-related span markers; seven opening and seven closing tags.
<spn/>	Speech present but not transcribable; 20 occurrences.
<unclear/>	Unclear audio; four occurrences.
<clear-throat/>	Throat clearing; six lowercase occurrences plus one <Clear-throat/> variant.

Tag spelling and punctuation are preserved from the source. Event counts above are counted in full transcripts, not inferred from the word arrays. Per the source documentation, event timings are interpolated around aligned lexical words rather than independently measured acoustic event boundaries. They should not be used as precision ground truth for breath or laughter onset detection.

Synchronizing the two speakers

Two speaker tracks in a session begin at slightly different times. The filename follows this pattern:

<room_start_epoch_ms>-<speaker_id>-hifi-audio-<capture_start_epoch_ms>.flac

To place both tracks on a common conversation timeline, subtract the earliest capture start in that session from each track's capture start and convert milliseconds to seconds. The later speaker's offsets in this release are 1.321, 2.125, 2.109, and 1.298 seconds for source conversations 01–04 respectively. Do not align both files to zero if studying overlap or turn-taking.

import json
from collections import defaultdict
from pathlib import Path
from huggingface_hub import hf_hub_download

repo_id = "OcularAIInc/AMERICAN-ENGLISH-TRANSCRIBED-HIFI-FULL-DUPLEX-TWO-SPEAKER-CONVERSATIONAL-DATASET-SAMPLE"
metadata_path = hf_hub_download(repo_id, "train/metadata.jsonl", repo_type="dataset")
rows = [json.loads(line) for line in Path(metadata_path).read_text().splitlines()]
sessions = defaultdict(list)
for row in rows:
    sessions[row["session_id"]].append(row)

def capture_ms(row):
    return int(Path(row["file_name"]).stem.rsplit("-", 1)[1])

for tracks in sessions.values():
    origin = min(capture_ms(row) for row in tracks)
    for row in tracks:
        offset = (capture_ms(row) - origin) / 1000
        # Add offset to this track's segment/word times for a shared timeline.
        print(row["session_id"], row["speaker_id"], offset)

Loading and listening

from datasets import load_dataset, Audio

repo_id = "OcularAIInc/AMERICAN-ENGLISH-TRANSCRIBED-HIFI-FULL-DUPLEX-TWO-SPEAKER-CONVERSATIONAL-DATASET-SAMPLE"
dataset = load_dataset(repo_id, split="train")
print(dataset.column_names)
print(dataset[0]["transcript"])
# To inspect audio references without requiring an audio decoder:
dataset = dataset.cast_column("audio", Audio(decode=False))
print(dataset[0]["audio"])

To listen to a specific track independently of the Datasets decoder:

import soundfile as sf
from IPython.display import Audio as NotebookAudio, display

row = rows[0]  # from the metadata example above
audio_path = hf_hub_download(repo_id, "train/" + row["file_name"], repo_type="dataset")
waveform, rate = sf.read(audio_path)
display(NotebookAudio(waveform, rate=rate))

Intended uses and evaluation guidance

This sample supports inspection of conversational ASR references, verbatim transcription workflows, alignment tooling, and paired-speaker timing analyses. Define normalization rules before comparing recognition outputs: whether to keep fillers and repetitions, how to treat punctuation, and how to exclude non-speech tags or unintelligible spans. Report those rules alongside any WER results.

The release does not establish any particular model's performance or training exposure. Once public, it should not be treated as a guaranteed unseen or contamination-free test set. Do not train and evaluate on the same recordings and interpret the result as held-out generalization. Use speaker-disjoint splits across larger collections; source speaker IDs may recur outside this sample.

Limitations

• Four conversations and eight speakers do not represent all American English speakers, accents, demographics, topics, microphones, or acoustic conditions.
• Discussion prompts are present. Natural interaction does not mean the recordings are entirely unprompted or representative of all production traffic.
• Four tracks have a missing age_range; missing values must not be filled with invented attributes.
• Two segments have empty words arrays. Segment text is still supplied, but those segments do not provide word-level timing coverage.
• One segment ends approximately 69 ms beyond its track's measured duration. Consumers requiring strict bounds should clamp or flag boundaries rather than silently assuming all timestamps are exact.
• Word alignment can be less reliable for short fillers and unstressed function words. Event boundaries are interpolated.
• Separate microphone capture is not a guarantee of zero acoustic bleed. Audio has not been reprocessed for this release.
• This publication check verified file/metadata correspondence and schema and screened transcript text for obvious email addresses, phone-number patterns, and URLs. That is not an exhaustive privacy audit or a claim of anonymity.

Consent, license, and attribution

Ocular confirmed that participant consent covers public redistribution of these recordings and transcripts. This release is made available under Creative Commons Attribution 4.0 International (CC BY 4.0). See LICENSE and the license text.

When redistributing or adapting the dataset, credit Ocular AI, link to this dataset and the license, and indicate changes. The license does not imply endorsement or eliminate independently applicable privacy or personality rights.

Suggested attribution: “Ocular AI, AMERICAN ENGLISH TRANSCRIBED HI-FI FULL-DUPLEX TWO-SPEAKER CONVERSATIONAL DATASET — SAMPLE (2026), CC BY 4.0.”

Contact

For other languages, accents, collection requirements, or annotation samples, contact Ocular or email founders@useocular.com.