---
license: mit
language:
- en
base_model:
- Qwen/Qwen3-ASR-1.7B
tags:
- speech
- recognition
- asr
---

This model was created for the [On Top of Pasketti: Children’s Speech Recognition Challenge - Word Track](https://www.drivendata.org/competitions/308/childrens-word-asr/) competition. It was trained on a large-scale dataset specifically designed for children's speech recognition.

Model is based on **Qwen/Qwen3-ASR-1.7B**.
* Local validation WER: 0.1005
* Private Leaderboard WER: 0.2037

Usage:

```python
import torch
from qwen_asr import Qwen3ASRModel

def get_dynamic_batches(items):
    total = len(items)
    i = 0

    while i < total:
        if items[i]['audio_duration_sec'] > 200:
            batch_size = 1
        else:
            batch_size = int(5000 / items[i]['audio_duration_sec']) + 1
            batch_size = min(batch_size, 64)

        yield items[i:i + batch_size]
        i += batch_size

model = Qwen3ASRModel.from_pretrained(
	"ZFTurbo/Qwen3-ASR-Children-Words",
	dtype=torch.bfloat16,
	device_map="cuda:0",
	max_inference_batch_size=64,
	max_new_tokens=-1,
)

with torch.inference_mode():
	for batch in get_dynamic_batches(items):
		paths = []
		languages = []
		max_new_tokens = 0
		for item in batch:
			path = str(data_dir / item["audio_path"])
			paths.append(path)
			languages.append("English")
			max_new_tokens = max(max_new_tokens, int(item["audio_duration_sec"] * 20))

		print(
			"Batch size:", len(batch),
			"Duration:", batch[0]["audio_duration_sec"],
			"Processed:", total,
			"Max tokens:", max_new_tokens,
		)
		cur_time = time.time()

		results = model.transcribe(
			audio=paths,
			language=languages,  # can also be set to None for automatic language detection
			return_time_stamps=False,
			max_new_tokens=max_new_tokens,
		)

	   predictions = {}
		for i, r in enumerate(results):
			predictions[i] = r.text

```

More usage examples: [https://github.com/ZFTurbo/Children-Speech-Recognition-Challenge-Solution](https://github.com/ZFTurbo/Children-Speech-Recognition-Challenge-Solution)