This variant is specifically trained to improve recognition of Japanese proper nouns, organization names, product names, service names, and kanji-heavy expressions that general-purpose ASR models tend to mistranscribe.

Intended Use

Designed for Japanese ASR tasks requiring accurate transcription of:

Proper nouns (people, places, organizations)
Difficult or uncommon kanji spellings
Company and product names
Mixed Japanese/English technical terminology

Example Normalizations

Product / Company Names

Spoken	Expected output
ジェミニ	`Gemini`
アンスロピック	`Anthropic`
ノートブックLM	`NotebookLM`

Japanese Numerals

Spoken	Expected output
一万二千三百四十五	`1万2345`
十人	`10人`

Math / Code

Spoken	Expected output
AイコールAプラス1	`A=A+1`

Model Details


Base model	`Qwen/Qwen3-ASR-1.7B`
Task	Automatic speech recognition
Language	Japanese (ja), English (en)
Fine-tuning focus	Japanese proper nouns & technical vocabulary

Quickstart

Install the runtime package:

pip install -U qwen-asr

Transcribe a local audio file:

python

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "neosophie/Qwen3-ASR-1.7B-JA",
    dtype=torch.bfloat16,
    device_map="cuda:0",
)

results = model.transcribe(audio="/path/to/audio.wav")

print(results[0].language)
print(results[0].text)

Acknowledgments

Qwen3-ASR-1.7B — base model by Alibaba Cloud

Built by Neosophie