Typhoon ASR Real-time: FastConformer-Transducer for Thai Automatic Speech Recognition

TL;DR

This work tackles the challenge of real-time Thai ASR by prioritizing data quality and architectural efficiency over sheer model size. It introduces Typhoon ASR Real-time, a 115M-parameter FastConformer-Transducer that achieves competitive accuracy with a $45\times$ computation reduction compared to large offline models, enabled by a rigorously normalized training pipeline and a consensus-based data curation strategy. The authors also present a two-stage curriculum for Isan dialect adaptation, and release the Typhoon ASR Benchmark to standardize evaluation and support reproducibility across Thai ASR research. Together, these contributions demonstrate that streaming Thai ASR with strong phonetic fidelity is feasible at a fraction of the computational cost, paving the way for practical on-device or low-resource deployments and more robust benchmarking in Thai language processing.

Abstract

Large encoder-decoder models like Whisper achieve strong offline transcription but remain impractical for streaming applications due to high latency. However, due to the accessibility of pre-trained checkpoints, the open Thai ASR landscape remains dominated by these offline architectures, leaving a critical gap in efficient streaming solutions. We present Typhoon ASR Real-time, a 115M-parameter FastConformer-Transducer model for low-latency Thai speech recognition. We demonstrate that rigorous text normalization can match the impact of model scaling: our compact model achieves a 45x reduction in computational cost compared to Whisper Large-v3 while delivering comparable accuracy. Our normalization pipeline resolves systemic ambiguities in Thai transcription --including context-dependent number verbalization and repetition markers (mai yamok) --creating consistent training targets. We further introduce a two-stage curriculum learning approach for Isan (north-eastern) dialect adaptation that preserves Central Thai performance. To address reproducibility challenges in Thai ASR, we release the Typhoon ASR Benchmark, a gold-standard human-labeled datasets with transcriptions following established Thai linguistic conventions, providing standardized evaluation protocols for the research community.

Figures (4)

• Figure 1: Pareto efficiency of Thai ASR models. The scatter plot illustrates the trade-off between computational cost (x-axis, GFLOPs per 30s audio segment on a logarithmic scale) and average Character Error Rate (y-axis) across three benchmarks. Our proposed streaming models, Typhoon ASR Realtime and Typhoon Isan ASR Realtime, occupy the optimal lower-left region (highlighted), demonstrating comparable accuracy to the offline state-of-the-art Pathumma-Whisper Large-v3 baseline while achieving an approximate 45$\times$ reduction in computational complexity.
• Figure 2: The Consensus Audio Transcription and Verification Pipeline. Raw audio is processed in parallel by three Whisper-Large models. A majority voting strategy selects consensus transcriptions, defaulting to Pathumma-Whisper-Large when no agreement exists. Complex transcriptions undergo human review, while clean outputs proceed directly to storage.
• Figure 3: Two-stage curriculum learning for dialect adaptation. Stage 1 employs low learning rate ($10^{-5}$) for gentle acoustic adaptation of the full model over 10 epochs. Stage 2 freezes the encoder (preserving acoustic stability) and employs high learning rate ($10^{-3}$) for rapid linguistic specialization of the decoder and joint network over 15 epochs.
• Figure 4: A/B testing results showing Win-Tie-Loss counts for Gemini 2.5 Pro against various competitor systems (N=1000 comparisons). The dashed red line marks the 500-count (50%) threshold.