Luel-ai/luel-conversational-speech-samples

Conversational Speech Samples (Luel)

License: All Rights Reserved. Proprietary. Access only for authorized parties; no redistribution or use without permission. See LICENSE.

A multilingual conversational speech dataset: two-speaker dialogue sessions across 8 languages. Each session includes combined audio, speaker-separated audio, and a JSON file containing transcripts with word-level timestamps, speaker diarization, and acoustic metadata.

---

Quick Stats

Metric	Value
Languages	9 (English, German, French, Spanish, Japanese, Hindi, Punjabi, Marwari, Urdu)
Total sessions	131
Total duration	68h 57m (248,278s)
Total utterances	26,600
Total words	698,641
Audio files	393 WAV
JSON files	131
Speakers per session	2
Sample rate	48 kHz
Total size on disk	~74.6 GB

Per-Language Breakdown

Language	Sessions	Duration	Utterances	Words	Audio Files
English	12	5h 35m	3,840	58,535	36
Hindi	9	4h 56m	695	48,787	27
Punjabi	14	6h 44m	316	67,670	42
Urdu	16	7h 22m	687	69,021	48
Marwari	17	7h 31m	4,096	88,566	51
German	13	6h 30m	1,571	64,752	39
French	13	5h 42m	1,213	39,630	39
Japanese	13	12h 30m	9,971	129,201	39
Spanish	9	4h 10m	1,035	39,156	27
Noisy English	15	7h 52m	3,176	93,323	45

Content Categories

Category	Sessions
Spontaneous Discussions	97
Formal Meetings	19

Acoustic Classification

Each session includes an objective acoustic spontaneity score (0-10) derived from overlap rates, turn-taking speed, backchannel frequency, F0 variability, and disfluency patterns.

Acoustic Category	Sessions
formal_interactive	48
spontaneous_natural	34
spontaneous_structured	18
formal_scripted	16

Combined classification (acoustic + content topic): 43/53 sessions (81.1%) classified as spontaneous.

---

Structure

Path	Description
en/american/	American English sessions (session_01 - session_09)
en/global/	Global English sessions (session_01 - session_03; non-American native English: African, French/Arabic)
de/	German sessions (13 topic-based folders)
fr/	French sessions (13 topic-based folders)
ja/	Japanese sessions (13 topic-based folders)
hi/	Hindi sessions (14 session folders)
pa/	Punjabi sessions (14 session folders)
mr/	Marwari sessions (17 session folders)
ur/	Urdu sessions (16 session folders)
es/	Spanish sessions (9 topic-based folders)
noisy_samples/	English noisy-background subset (acoustic_category: noisy_background) — see Noisy Samples Subset below

Each session folder contains exactly four files:

{name}.json              Metadata, transcript, word-level timestamps, acoustic metadata
{name}.wav               Combined audio (both speakers)
{name}_speaker1.wav      Speaker 1 isolated audio
{name}_speaker2.wav      Speaker 2 isolated audio

---

Noisy Samples Subset

The noisy_samples/ directory holds English conversational sessions with verified acoustic background noise (crowd voices, traffic, animals, static, indoor ambience, etc.) but otherwise clear primary speech. Sessions in this subset are intended for noise-robust ASR training, acoustic-event detection, and downstream robustness benchmarking.

Each session in noisy_samples/ follows the same four-file layout as the main dataset ({name}.json, {name}.wav, {name}_speaker1.wav, {name}_speaker2.wav) plus an additional noise_classification block inside the JSON metadata.

Structure

noisy_samples/
└── en/                  English noisy sessions

Acoustic Distinction

Sessions in noisy_samples/ carry acoustic_category: "noisy_background" (distinct from the spontaneous/formal categories used in the main dataset). Primary speech is fully transcribable; the noise is a backdrop rather than the foreground.

Additional JSON field: noise_classification

In addition to all the standard fields documented in the JSON Schema section, each noisy-sample JSON contains:

"noise_classification": {
  "noise_types": ["crowd_voices", "traffic_indoor", "animal_dog", "static_noise", "background_music", "..."],
  "noise_severity": "light" | "moderate" | "heavy",
  "tier2_evidence": ["BG_VOICES", "BG_TRAFFIC", "BG_INDOOR", "BG_OUTDOOR", "STATIC", "ANIMAL", "MUSIC"],
  "labeler_notes": "Free-text description of the noise from the QA labeler",
  "labeled_at": "ISO-8601 timestamp",
  "labeled_channels": ["host", "guest", "both"]
}

Field	Type	Description
noise_types	string[]	Specific noise categories present in the session (free-text taxonomy)
noise_severity	enum	Subjective severity: light / moderate / heavy
tier2_evidence	string[]	Standardized acoustic-event tags supporting the classification
labeler_notes	string	Human-readable description (e.g., "People (about 3) talking in the background in guest channel")
labeled_at	string	When the noise label was applied
labeled_channels	string[]	Which track(s) carry the noise: host, guest, or both

Use cases

• Training noise-robust speech recognizers
• Acoustic-event detection (which noise types co-occur)
• Active-noise-cancellation evaluation
• Robustness ablations for downstream models (compare clean vs noisy-background performance)

---

JSON Schema

Each session JSON contains these top-level fields:

Field	Type	Description
UUID	string	Unique session identifier
audioURL	string	Relative path to combined audio
speaker1_audioURL	string	Relative path to speaker 1 audio
speaker2_audioURL	string	Relative path to speaker 2 audio
name	string	Session name
durationSeconds	float	Total duration in seconds
language	string	Language name
category	string	spontaneous_discussions or formal_meetings
topic	string	Conversation topic
numSpeakers	int	Always 2
sampleRate	int	Always 48000
metadata	object	Recording context (language, country, region, environment, device, topic)
transcript	string	Full concatenated transcript
speakerTranscript	array	Speaker-attributed utterances with word-level timing
acoustic_metadata	object	Objective acoustic descriptors and classification
noise_classification	object	(only present in noisy_samples/ sessions) Verified background-noise labels — see Noisy Samples Subset

speakerTranscript

Array of utterance objects, each containing:

{
  "speaker": "Speaker1",
  "text": "How have you been?",
  "startTime": 0.24,
  "endTime": 1.58,
  "words": [
    { "text": "How", "startTime": 0.24, "endTime": 0.4 },
    { "text": "have", "startTime": 0.42, "endTime": 0.54 },
    { "text": "you", "startTime": 0.56, "endTime": 0.64 },
    { "text": "been?", "startTime": 0.66, "endTime": 1.58 }
  ]
}

acoustic_metadata

Nested object with these sub-groups:

Sub-group	Key Fields
loudness	integrated_lufs, lra_lu, true_peak_dbfs, short-term stats
loudness_consistency	30-second segment mean/std/range LUFS
snr	combined_snr_db, quality_tier
clipping	clip_ratio_pct, clipping_event_count, max_consecutive_clipped
vad	speech_rate, silence_rate, overlap_rate, per-speaker rates
turn_taking	turns_per_minute, median_turn_duration_sec, backchannel_count, speaking_rate_std_wpm
noise	noise_rms_db, spectral_centroid_mean_hz, spectral_flatness_mean
environment	reverb_category, c50_db
bandwidth	type (fullband/wideband/narrowband), narrowband_detected
silence	ratio, segments, mean_duration_sec
quality	mos_overall, mos_signal, mos_background, mos_speech_quality
per_speaker	Per-speaker LUFS, SNR, clipping, F0 (mean/std/range/variability), turns, WPM
classification	acoustic_category, spontaneity_score, content_category, is_spontaneous_combined

---

Access

This repository is gated. You must accept the conditions to access files and content. For access requests, contact Luel (https://luel.ai).