typhoon-asr-qwen-0.6b-ctx

Thai ASR with prompt-based contextual biasing: a LoRA fine-tune (merged, 0.80B params) of Qwen/Qwen3-ASR-0.6B-hf that accepts a list of names/terms in the system prompt, boosts the ones that are actually spoken, writes them with the listed spelling, and ignores the rest (distractor false alarm rate ≤.0004; non-speech probe 5/5). The economical tier: 2.6× smaller than the 1.7b for ~.09 recall.

Results

Plain transcription (CER%, same protocol: gigaspeech2 bench in-domain / TVSpeech out-of-domain 30-s talk-show):

† single-digit repetition-loop utterances dominate (0.6b: one utt; Thonburian: three, excluding them 8.39).

Contextual biasing — vs every route we measured

Same benchmark (wayu-ai/thai-contextasr-bench), same frozen scorer. Cells = CER% ↓ / entity recall ↑. Whisper models take the list as a decoder prompt (zero-shot); the two typhoon-qwen models are LoRA fine-tunes; typhoon-asr-realtime-nemo-ctc spots-and-splices at decode time.

N/A: typhoon-whisper-large-v3's prompt pathway collapses on any list (echo loops, CER 200+); the whisper fine-tunes cannot take lists beyond ~10–15 entries because Whisper's 448-token decoder window is shared between the prompt and the transcript. Only the typhoon-qwen pair and the typhoon-asr-realtime-nemo-ctc spotter survive large lists.

Usage

import torch, soundfile as sf
from transformers import AutoModelForMultimodalLM, AutoProcessor  # transformers>=5.13

mid = "typhoon-ai/typhoon-asr-qwen-0.6b-ctx"
processor = AutoProcessor.from_pretrained(mid)
model = AutoModelForMultimodalLM.from_pretrained(mid, dtype=torch.bfloat16,
                                                 device_map="cuda").eval()
audio, sr = sf.read("clip.wav", dtype="float32")   # 16 kHz mono

def transcribe(system_text):
    msgs = [{"role": "system", "content": [{"type": "text", "text": system_text}]},
            {"role": "user",   "content": [{"type": "audio", "audio": audio}]}]
    inputs = processor.apply_chat_template([msgs], tokenize=True,
        add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)
    out = model.generate(**inputs, max_new_tokens=440)
    out = out[:, inputs["input_ids"].shape[1]:]
    return processor.decode(out, return_format="transcription_only")[0].strip()

print(transcribe("Thai"))                                   # plain ASR
print(transcribe("วโรรส Facebook ก้าวไกล\nThai"))            # with a bias list

Prompt format is strict: the system message is the space-joined bias list, then Thai on its own line — or just Thai with no list. Recall decays gracefully with list length; ≤50 entries is the sweet spot.

License & attribution

Apache-2.0 — LoRA-merged weights of Qwen/Qwen3-ASR-0.6B-hf (Apache-2.0). Trained on ~10,800 hours of Thai speech (GigaSpeech2-th); see the Typhoon ASR Real-time technical report.

Acknowledgements

We thank Kunat Pipatanakul and Potsawee Manakul (wayu-ai) for their valuable feedback.