In the past, speech recognition went blind the moment it hit a dialect: as soon as Shanghainese, Hokkien, or Cantonese came out, the transcript often turned to gibberish or came back blank. Fun-ASR1.5, the latest release from Alibaba's Tongyi Lab, changes that. It is an end-to-end speech recognition foundation model in which a single model covers 30 languages and the seven major Chinese dialect groups, with the character error rate in dialect scenarios down 56.2% versus the previous generation.

What Is Fun-ASR1.5

Fun-ASR1.5 is the new-generation version of Alibaba Qwen's end-to-end speech recognition foundation model, trained on hundreds of thousands of hours of real speech data. It doesn't just "hear clearly" — it "understands what it hears", intelligently inserting punctuation and automatically normalizing spoken expressions, so the output text reads close to written quality.

Core Capabilities

Dialect recognition: from "unintelligible" to "industrial-grade usable"

It covers the seven traditional Chinese dialect groups (Mandarin, Wu, Xiang, Gan, Hakka, Min, and Yue) and is deeply adapted to 20+ regional accents, including Sichuanese, Hokkien, the Changsha dialect, the Suzhou dialect, and Cantonese.

Real-world test data:

• Accuracy above 90% on 5 dialects
• Accuracy above 80% on 15 dialects
• Character error rate (CER) in typical dialect scenarios down 56.2% relative
• Beats Seed-ASR and Tencent-ASR on 13 dialects

30 languages + free cross-language switching

It supports 30 languages including Chinese, English, Japanese, Korean, French, German, and Spanish. In mixed-language conversations, it automatically detects and switches without any preset language tags.

For example, in a conversation mixing English and Japanese, the model accurately distinguishes and transcribes both languages.

Dedicated optimization for classical poetry

It received dedicated training for classical Chinese poetry recognition, with an audio-to-text aligned corpus of poems spanning from the pre-Qin era to modern times. Internal evaluation puts character-level accuracy at 97%.

Smart punctuation and text normalization

• Punctuation prediction: Automatically inserts commas, periods, question marks, and more based on semantics — "how is the weather today i want to go out for a walk" -> "How is the weather today? I want to go out for a walk, but I'm afraid it might rain."
• Text normalization: Automatically converts spoken expressions — "three thousand five hundred and sixty-two" -> "3562", "fifty-eight thousand yuan" -> "58000 RMB"

Quick Start

Try it online

• ModelScope community demo: https://modelscope.cn/studios/iic/FunAudio-ASR
• Alibaba Cloud Bailian platform: https://bailian.console.aliyun.com/cn-beijing?tab=model#/efm/model_experience_center/voice?modelId=fun-asr

API call example

Call it through the Alibaba Cloud Bailian platform — a few lines of code and you're integrated:

from http import HTTPStatus
from dashscope.audio.asr import Transcription
import dashscope
import os
import json

dashscope.base_http_api_url = 'https://dashscope.aliyuncs.com/api/v1'
dashscope.api_key = os.getenv("DASHSCOPE_API_KEY")

task_response = Transcription.async_call(
    model='fun-asr',
    file_urls='https://example.com/your-audio.wav'
)

transcribe_response = Transcription.wait(task=task_response.output.task_id)
if transcribe_response.status_code == HTTPStatus.OK:
    print(json.dumps(transcribe_response.output, indent=4, ensure_ascii=False))

Application Scenarios

• Meeting minutes: Smart punctuation + text normalization dramatically reduce proofreading work afterward
• News interviews: Accurate transcription even in noisy environments, with cross-language support
• Education: Transcribing recitations of classical poetry, supporting Chinese classics education and intangible cultural heritage preservation
• Government services: Transcribing dialect hotlines, covering county-level and rural government scenarios
• Cross-border collaboration: Multilingual meeting records with no preset language tags
• Content production: Subtitle generation for short videos, podcast-to-text transcription