LOTUSDIS: Thai Far-Field Meeting ASR Corpus
- LOTUSDIS is a Thai far-field meeting speech corpus featuring spontaneous, overlapping conversations captured by diverse off-the-shelf devices.
- It provides 114 hours of multi-channel recordings with standardized train, dev, and test splits to benchmark realistic ASR performance.
- Empirical findings reveal significant zero-shot performance gaps that are mitigated by distance-diverse fine-tuning and controlled augmentation techniques.
Searching arXiv for LOTUSDIS and closely related work to ground the article in current papers. LOTUSDIS is a publicly released Thai far-field meeting corpus created to advance robust conversational ASR in realistic, single-channel conditions. It is designed around spontaneous multi-speaker interaction recorded simultaneously by diverse, off-the-shelf devices at multiple distances, rather than microphone arrays, and is accompanied by reproducible training and evaluation scripts plus a benchmark based on Whisper variants in zero-shot and fine-tuned settings. The corpus targets a resource gap in Thai speech technology: distant, overlapping conversational speech under reverberation, noise, and device coloration, with standard train, dev, and test splits and an open CC-BY-SA 4.0 release (Tipaksorn et al., 23 Sep 2025).
1. Research objective and corpus identity
LOTUSDIS was introduced to address a specific mismatch between widely used Thai speech resources and the conditions encountered in real meetings. Existing Thai corpora are described as emphasizing near-field read speech or broadcast audio and often carrying restrictive licenses, whereas LOTUSDIS targets spontaneous, unscripted dialogue with frequent overlap, large microphone-distance variation, and realistic room acoustics. The corpus is therefore intended for robust conversational ASR rather than clean, single-speaker transcription, and its design explicitly prioritizes single-channel realism over array-based capture or beamforming assumptions (Tipaksorn et al., 23 Sep 2025).
Its core contributions are threefold. First, it provides an open corpus comprising 114 hours total across microphones. Second, it includes reproducible baselines: fine-tuning recipes for Whisper, standardized decoding and scoring, and per-device benchmarks. Third, it supplies empirical evidence that off-the-shelf models exhibit a strong domain mismatch on Thai far-field speech, while distance-diverse fine-tuning substantially improves robustness, especially on the most distant microphones. This framing places LOTUSDIS at the intersection of Thai ASR, far-field speech recognition, overlapped speech benchmarking, and device-robust meeting transcription.
2. Corpus composition and recording environment
The corpus contains 90 sessions total, each 15–20 minutes long, with three participants per session engaged in spontaneous, topic-prompted conversation. The language is Thai in standard conversational form, with annotators marking dialectal uncertainty when present. Across microphones, the corpus totals 114 hours of multi-channel audio derived from approximately 20 hours of unique meeting time. The split statistics are detailed and balanced enough to support standardized benchmarking: the training split contains 69 sessions, 17:37 of unique meeting duration, 88:07 across selected microphones, 74 unique participants, and 120,245 utterances across 40 topics; the development split contains 10 sessions, 2:33 of unique meeting duration, 12:49 across selected microphones, 12 participants, and 13,090 utterances across 10 topics; the test split contains 11 sessions, 2:39 of unique meeting duration, 13:17 across selected microphones, 12 participants, and 27,580 utterances across 11 topics (Tipaksorn et al., 23 Sep 2025).
The participant pool comprises 86 unique participants aged 19–48, with mean age 27. Overlap is central rather than incidental: roughly one-third of speech is overlapped, with split-specific overlap ratios of 31.5% for train, 40.3% for dev, and 29.0% for test. This makes LOTUSDIS suitable not only for standard WER evaluation but also for overlap-sensitive analyses.
The recording setup emphasizes distance diversity within a single scene. Speech was captured simultaneously by nine independent single-channel devices spanning six microphone types. The named benchmark categories are three lavalier microphones and three table-mounted condensers at 12–15 cm from each speaker’s mouth, plus one tabletop “JBL” device at 2 m and two Bluetooth speakerphones at 3 m and 10 m. The room was a furnished office measuring m, with HVAC, refrigerator, and water cooler left on to create a stationary, realistic noise floor. Far-field devices exhibited lower SNR, stronger reverberation, and high-frequency roll-off, while near-field channels provided high direct-to-reverberant ratios. Alignment was performed via slate pulses and cross-correlation to achieve sub-sample alignment and minimal drift. Specific sampling rate, bit depth, file formats, and RT60 measurements are not reported in the paper; the repository is indicated as the source for exact audio-format details.
3. Annotation design and metadata structure
LOTUSDIS uses a two-stage labeling process. In the first stage, three trained annotators segment and transcribe the recordings with unified guidelines covering tokenization, code-switching, and non-lexical events. In the second stage, a senior annotator reviews each session to fix boundaries and resolve discrepancies. This workflow is intended to preserve consistency under spontaneous and frequently overlapped dialogue conditions (Tipaksorn et al., 23 Sep 2025).
The released labels include utterance-level transcripts with speaker IDs, explicit overlap masks, and time-stamped utterances tied to speaker IDs and channel. Overlapped spans are indicated by concatenating speaker IDs with an ampersand. The tag set includes <n> for noise, <sil> for silences greater than 300 ms, <unk> for unintelligible speech, and <td> for dialectal uncertainty. The combination of speaker identity, overlap annotation, and channel association makes the corpus appropriate for controlled evaluations of overlap-aware ASR and related tasks.
This annotation design has methodological implications. Because overlap is explicitly marked rather than only implicit in the waveform, LOTUSDIS supports analyses that separate single-speaker from overlapped spans. A plausible implication is that it can serve as a benchmark not only for transcription quality but also for error-mode analysis under overlap, distance, and device mismatch.
4. Baseline system and evaluation protocol
The accompanying baseline system standardizes decoding, text normalization, and scoring, and benchmarks several Whisper variants under both zero-shot and fine-tuned conditions. The evaluated models are Whisper-large-v3, Pathumma-whisper-th-large-v3, Biodatlab/whisper-th-large-v3-combined, and Monsoon-whisper-medium-gigaspeech2. For scoring, Thai word tokenization is performed with the PyThaiNLP newmm segmenter so that WER is computed consistently across conditions. The same decoding and scoring pipeline is used across models, although beam settings are not enumerated in the paper (Tipaksorn et al., 23 Sep 2025).
The fine-tuning recipe uses the LOTUSDIS training split, typically in an “All Mic” configuration, on a single NVIDIA H200 GPU for 5 epochs. Optimizer, learning rate, and batch size are not specified in the paper. Additional experiments study single-microphone training with SpecAugment and with reverberation and additive-noise augmentation derived from OpenSLR-28 RIRs plus HVAC and conversational noise.
The principal evaluation metric is WER on Thai text after newmm segmentation, reported per microphone as well as via near-field, far-field, and overall macro averages. The paper also gives CER, SNR, and RT60 definitions. The reported formulas are
and
with the latter computed over characters. SNR is defined as
although numerical SNR values are not reported. This evaluation design places emphasis on per-device robustness and on the near-field/far-field generalization gap rather than only aggregate accuracy.
5. Empirical findings and technical implications
The main empirical result is a strong zero-shot domain mismatch for Thai far-field meeting speech. Off-the-shelf Whisper-large-v3 achieves an overall WER of 79.84% on the LOTUSDIS test set. Among Thai-adapted off-the-shelf models, Pathumma shows the clearest summary of the near–far disparity: zero-shot macro averages are 36.99% near-field, 81.57% far-field, and 64.32% overall. The paper characterizes performance as degrading monotonically with distance in zero-shot evaluations, indicating that pre-training data are biased toward higher-SNR, near-field conditions (Tipaksorn et al., 23 Sep 2025).
Fine-tuning on LOTUSDIS materially changes this regime. The Pathumma-whisper-th-large-v3 baseline fine-tuned on all microphones reaches 38.33% overall WER, 21.59% near-field macro WER, and 49.54% far-field macro WER, with per-device WERs of 26.42% on JBL, 58.15% on BT3m, and 64.04% on BT10m. Whisper-large-v3 fine-tuned on LOTUSDIS reaches 39.05% overall, 22.01% near-field macro, and 50.51% far-field macro. For Pathumma specifically, the overall WER is reduced from 64.3% to 38.3%, and the far-field WER from 81.6% to 49.5%. The gains are described as especially large at the most distant microphones.
The paper also reports negative results that are methodologically important. Front-end dereverberation and denoising are not universally beneficial in this mixed-distance setting: WPE increases near-field and far-field WER to 35.92% and 56.12%, respectively, and MMSE-LSA increases them to 24.92% and 54.55%, relative to 21.59% and 49.54% without such front-end processing. This suggests that blanket enhancement can damage already strong near-field channels while failing to deliver compensatory gains on distant channels.
Single-microphone training reveals an additional failure mode. Training only on near-field condenser data yields strong in-domain performance, including condenser WER of 19.26%, but collapses on far-field devices, with BT3m at 97.95%, BT10m at 113.65%, macro far-field at 79.54%, and overall at 50.12%. Augmentation mitigates but does not eliminate this gap: condenser plus SpecAugment yields 49.11% overall and 67.15% far-field, while condenser plus reverberation and noise yields 45.86% overall and 65.39% far-field. The implication drawn in the paper is that distance-diverse real recordings remain more effective than single-device overfitting, although physically motivated augmentation is useful when multi-device data are unavailable.
Overlap analysis further sharpens the corpus’s value. Overlapped spans shift WER distributions higher for all microphones, most severely at BT3m and BT10m. The observed errors include more deletions and short-token drops, while substitutions are frequent on tone-bearing syllables, which the paper interprets as indicating masking of critical tonal cues. This is particularly consequential in Thai, where tone-bearing syllables carry lexical distinctions.
6. Reproducibility, research uses, limitations, and nomenclature
LOTUSDIS is released under CC-BY-SA 4.0, with download, documentation, and baseline scripts available through the repository maintained by Pattara Tipaksorn, Sumonmas Thatphithakkul, Vataya Chunwijitra, and Kwanchiva Thangthai of NECTEC’s Speech and Text Understanding Research Team (Tipaksorn et al., 23 Sep 2025). The intended use cases include robust far-field Thai ASR, meeting transcription in single-channel deployments, overlap-aware ASR benchmarking, distance-aware training and evaluation, domain adaptation for Thai conversational speech across devices and rooms, and multi-device modeling without arrays.
The paper also states several limitations. LOTUSDIS is Thai-only, single-channel, and recorded in one office environment, so broader multilingual coverage, multiple rooms, and wider RT60 diversity would strengthen generalization. Although 114 hours across devices is substantial for Thai far-field meetings, the authors note that more sessions, speakers, and dialectal breadth would be helpful. Overlap remains difficult and may benefit from dedicated separation annotations and multi-speaker labels beyond masks. These constraints are best read as scope conditions rather than defects: the corpus is deliberately focused on robust Thai far-field meeting ASR under realistic, non-array conditions.
LOTUSDIS should also be distinguished from unrelated arXiv works that use the LOTUS or Lotus name in other domains, including a learned cutoff-table method for diversified nearest-neighbor retrieval, a stellar-parameter derivation tool, a low-cost ultraviolet spectrograph, a looped-transformer latent reasoning method, and an edge-device DVFS controller for two-stage detectors (Matsui, 5 Jun 2025, Li et al., 2022, Steele et al., 2016, Fan et al., 30 Jun 2026, Gong et al., 2024). Within speech and language technology, however, LOTUSDIS specifically denotes the Thai far-field meeting corpus and benchmark introduced for robust conversational ASR.