Typhoon Whisper Turbo

Typhoon Whisper Turbo is a state-of-the-art Thai Automatic Speech Recognition (ASR) model fine-tuned on the OpenAI Whisper Large v3 Turbo architecture. It is engineered to deliver a balance of high accuracy and low latency for offline transcription, significantly outperforming standard Whisper models in throughput while maintaining robust performance on Thai speech.

The model was presented in the paper Typhoon ASR Real-time: FastConformer-Transducer for Thai Automatic Speech Recognition.

Project Page: opentyphoon.ai
GitHub Repository: scb-10x/typhoon-asr

The model was trained on approximately 11,000 hours of Thai audio, curated and normalized using the SCB 10X Typhoon data pipeline to ensure consistent handling of Thai numbers, repetition markers, and context-dependent ambiguities.

Model Overview

Architecture: Whisper Large v3 Turbo (4 decoder layers vs 32 in standard Large v3)
Language: Thai
Dataset: ~11,000 hours of normalized Thai speech (Gigaspeech2, CommonVoice, Internal Curated Public Media)
Task: Automatic Speech Recognition (ASR)
License: MIT (inherited from OpenAI Whisper)

Performance

Typhoon Whisper Turbo achieves competitive performance against large-scale offline models on Thai speech recognition tasks.

Note: Lower CER (Character Error Rate) is better. Results on Gigaspeech2 (Clean/Academic), TVSpeech (Noisy/In-the-wild), and Google Fleurs (Thai) testset.

Usage

You can use this model directly with the Hugging Face transformers library.

Installation

pip install transformers torch accelerate

Example Code

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32

model_id = "scb10x/typhoon-whisper-turbo"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=448,
    chunk_length_s=30,
    batch_size=16,
    return_timestamps=True,
    torch_dtype=torch_dtype,
    device=device,
)

# Transcribe
result = pipe("path_to_audio.wav", generate_kwargs={"language": "thai"})
print(result["text"])

Training Data

The model was trained on approximately 11,000 hours of Thai audio data, including:

Gigaspeech2: Clean, academic-style speech
CommonVoice: Crowd-sourced diverse speech samples
Internal Curated Public Media: Proprietary datasets curated by Typhoon Team, SCB 10X

All data was normalized using the Typhoon data pipeline to ensure:

Consistent handling of Thai numbers
Proper treatment of repetition markers
Resolution of context-dependent ambiguities

Model Architecture

Typhoon Whisper Turbo is based on OpenAI Whisper Large v3 Turbo, achieving performance comparable to state-of-the-art models while being significantly more efficient:

Reduced model size: Only 4 decoder layers (compared to 32 in standard Large v3)
Lower memory footprint: Requires substantially less GPU/CPU memory during inference
Maintained accuracy: Competitive performance with minimal trade-off despite the compact architecture

Limitations

The model is optimized specifically for Thai language speech recognition
Performance may vary on dialects or accents not well-represented in the training data
Best suited for offline transcription rather than real-time streaming applications

License

This model is released under the MIT License, inherited from OpenAI Whisper.

Citation

If you use this model in your research or application, please cite our technical report:

@misc{warit2026typhoonasrr,
      title={Typhoon ASR Real-time: FastConformer-Transducer for Thai Automatic Speech Recognition}, 
      author={Warit Sirichotedumrong and Adisai Na-Thalang and Potsawee Manakul and Pittawat Taveekitworachai and Sittipong Sripaisarnmongkol and Kunat Pipatanakul},
      year={2026},
      eprint={2601.13044},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.13044}, 
}