Advancing Speech Language Models by Scaling Supervised Fine-Tuning with Over 60,000 Hours of Synthetic Speech Dialogue Data

Shuaijiang Zhao, Tingwei Guo, Bajian Xiang, Tongtang Wan, Qiang Niu, Wei Zou, Xiangang Li
Beike Inc., Beijing, China

Abstract

The GPT-4o represents a significant milestone in enabling real-time interaction with large language models (LLMs) through speech, its remarkable low latency and high fluency not only capture attention but also stimulate research interest in the field. This real-time speech interaction is particularly valuable in scenarios requiring rapid feedback and immediate responses, dramatically enhancing user experience. However, there is a notable lack of research focused on real-time large speech language models, particularly for Chinese. In this work, we present KE-Omni, a seamless large speech language model built upon Ke-SpeechChat, a large-scale high-quality synthesized speech interaction dataset consisting of 7 million Chinese and English conversations, featuring 42,002 speakers, and totaling over 60,000 hours, This contributes significantly to the advancement of research and development in this field. Furthermore, we plan to release the code and models after a thorough risk assessment.

KE-Omni Model Architecture

KE-Omni comprises three main components: a speech encoder, a large language model(LLM), and a speech decoder.

am-omni architecture

Ke-SpeechChat Dataset Construction

Textual Dialogue Construction

textual dialogue construction

Speech Dialogue Construction

speech dialogue construction

Ke-SpeechChat Dataset Details

Statistics of Ke-SpeechChat

Statistics of KESpeechChat

Subsets of Ke-SpeechChat

Subsets of KESpeechChat

Ke-SpeechChat Dialogue Demos

Chinese Dialogues

User-male, Agent-male

User: 北京一年的平均气温是多少度啊?

Agent: 北京一年的平均气温大概在十二度左右,不过四季温差挺大的,冬天会冷很多,夏天又热一些。

User-male, Agent-female

User: 人工智能现在都用在哪些地方了?

Agent: 人工智能的应用范围非常广泛,比如智能家居、自动驾驶汽车、医疗诊断、金融分析、教育辅导、娱乐推荐系统等等,几乎涵盖了我们生活的方方面面。

User-female, Agent-male

User: 人类的平均体温一般是多少度呀?

Agent: 人类的平均体温通常是在三十六点五到三十七点五摄氏度之间,这个范围内的体温都算是正常的。

User-female, Agent-female

User: 这句话:读万卷书,行万里路是什么意思呢?

Agent: 这句话的意思是说读书和实践都很重要。读万卷书指的是要多读书,增长知识;行万里路则是指要多出去走走,增长见识。结合起来就是理论和实践都要兼顾,这样才能更好地成长。

English Dialogues

User-male, Agent-male

User: What kinds of birds and trees would you usually see in a peaceful forest clearing?

Agent: In a peaceful forest clearing, you might see a variety of birds like robins, sparrows, and woodpeckers. As for trees, you could expect to find oaks, maples, and pines, which provide a beautiful and serene backdrop.

User-male, Agent-female

User: What are some things I should think about before agreeing to speak at a conference?

Agent: When considering a speaking engagement, think about the alignment of the conference theme with your expertise, the audience demographics, the potential networking opportunities, the time commitment required, and any travel or financial arrangements. Also, consider how this opportunity fits into your career goals and whether it can help you reach a wider audience or gain new insights.

User-female, Agent-male

User: Could you tell me what the capital of France is?

Agent: The capital of France is Paris.

User-female, Agent-female

User: What's the best way to store different types of garden produce to keep them fresh for as long as possible?

Agent: To keep your garden produce fresh, it's best to store each type according to its specific needs. For example, leafy greens like lettuce and spinach should be kept in a cool, humid environment, while root vegetables like carrots and potatoes prefer a cooler, drier place. Berries and stone fruits do well in the refrigerator, but tomatoes should be stored at room temperature. Make sure to avoid storing ethylene-producing fruits like apples and bananas with sensitive vegetables, as this can cause premature ripening.

Experiment Results

Results

Performances of speech language models on Ke-SpeechChat chat-test.
Models Data Chinese English
S2TIF Modal Align Quality S2TIF Modal Align Quality
Content Style Length CER UTMOS Content Style Length WER UTMOS
Qwen2-Audio - 3.32 3.09 242.74 - - 2.39 2.85 115.58 - -
LLaMA-Omni - - 2.85 3.70 46.80 9.27 3.95
SpeechGPT - - 2.86 3.37 57.00 56.34 3.90
KE-Omni(ours) XS 3.04 3.79 90.91 29.35 2.74 2.79 3.61 72.77 36.10 3.06
S 3.47 4.04 90.70 7.61 3.26 2.96 3.69 68.83 8.81 3.90
M 3.89 4.24 89.29 5.19 3.39 3.45 3.96 70.02 4.54 4.26
L 3.95 4.28 89.54 5.03 3.39 3.57 4.00 69.00 3.92 4.29
XL 4.12 4.34 88.10 5.16 3.43 3.61 4.00 68.70 4.29 4.30
Results of VoiceBench. For detailed results, please visit [VoiceBench Leadboard].KE-Omni-v1.6 trained on a smaller but higher-quality dataset filtered by LLM.
Model AlpacaEval CommonEval SD-QA MMSU OpenBookQA IFEval AdvBench Overall
KE-Omni-v1.6 4.12 3.53 45.03 33.12 48.13 18.41 100.0 56.81
GLM-4-Voice 3.97 3.42 36.98 39.75 53.41 25.92 88.08 55.99
DiVA 3.67 3.54 57.05 25.76 25.49 39.15 98.27 55.70
Qwen2-Audio 3.74 3.43 35.71 35.72 49.45 26.33 96.73 55.35
KE-Omni-v1.5 3.82 3.20 31.20 32.27 58.46 15.00 100.00 53.90
LLaMA-Omni 3.70 3.46 39.69 25.93 27.47 14.87 11.35 37.51
VITA 3.38 2.15 27.94 25.70 29.01 22.82 26.73 34.68
Mini-Omni2 2.32 2.18 9.31 24.27 26.59 11.56 57.50 31.32
Mini-Omni 1.95 2.02 13.92 24.69 26.59 13.58 37.12 27.90
Moshi 2.01 1.60 15.64 24.04 25.93 10.12 44.23 27.47

Case Study: English Dialogues

Ground Truth

User: What kind of job would suit someone who likes solving problems and interacting with people?

Agent: A career in fields like customer service, human resources, or project management could be a great fit. These roles often require strong problem-solving skills and regular interaction with people, which can be both challenging and rewarding.

SpeechGPT

User: What kind of job would suit someone who likes solving problems and interacting with people?

Agent: You may be referring to a job in customer service. This type of job involves interacting with customers and solving their problems. It requires excellent communication skills and the ability to remain calm under pressure.

LLaMA-Omni

User: What kind of job would suit someone who likes solving problems and interacting with people?

Agent: A job that suits someone who likes solving problems and interacting with people is a consultant or a project manager.

KE-Omni

User: What kind of job would suit someone who likes solving problems and interacting with people?

Agent: A job that combines problem-solving skills with people interaction could be a customer service representative or a project manager. These roles often require creative thinking to resolve conflicts and communicate effectively with various clients and team members. Another great option might be a salesperson, where you can build strong relationships and help clients find the right solutions.

Case Study: Chinese Dialogues

Ground Truth

User: 为啥青藏高原被叫做世界的第三极呢?

Agent: 青藏高原被称为世界第三极,主要是因为它拥有除南北极之外最大的冰川群,还有丰富的冰雪资源,平均海拔四千米以上,对全球气候和生态环境有着重要影响。

KE-Omni

User: 为啥青藏高原被叫做世界的第三极呢?

Agent: 青藏高原被称为世界的第三极,主要是因为它拥有地球上最大的高海拔地区,平均海拔超过四千米,拥有独特的生态系统和生物种类,科学研究价值非常高。

Ground Truth

User: 日光浴和徒步旅行都挺不错的,你觉得哪种更适合想放松心情的人呢?

Agent: 嗯,这得看个人喜好。不过,要是想彻底放松心情的话,我推荐日光浴,躺在温暖的阳光下,感受微风拂面,闭上眼睛听海浪声,整个人都会变得轻松起来。当然了,如果你喜欢在大自然中活动,徒步旅行也是个不错的选择,可以边走边欣赏美景,同样能让人感到愉悦。

KE-Omni

User: 日光浴和徒步旅行都挺不错的,你觉得哪种更适合想放松心情的人呢?

Agent: 嗯,这得看个人喜好。不过一般来说,日光浴可以让人感到温暖舒适,有助于放松心情。而徒步旅行则能让人亲近大自然,呼吸新鲜空气,也是个不错的选择。两种方式各有千秋,看您喜欢哪种了。

Ground Truth

User: 三十二岁的程序员通常会面临哪些工作上的挑战呢?

Agent: 三十二岁的程序员可能会遇到技术更新快,需要不断学习新技能的挑战。还有就是职业发展路径的选择,比如是继续深耕技术还是转向管理岗位。此外,工作与生活的平衡也是一个不小的难题。

KE-Omni

User: 三十二岁的程序员通常会面临哪些工作上的挑战呢?

Agent: 三十二岁的程序员可能会遇到职业发展的瓶颈,比如技术更新快,需要不断学习新技能,还有就是职位晋升的压力,可能还要面对工作和生活的平衡问题。

BibTeX

@article{zhao2024keomni,
  author    = {Zhao, Shuajiang and Guo, Tingwei and Xiang, Bajian and Wan, Tongtang and Niu, Qiang and Zou, Wei and Li, Xiangang},
  title     = {Advancing Speech Language Models by Scaling Supervised Fine-Tuning with Over 60,000 Hours of Synthetic Speech Dialogue Data},
  year      = {2024},
}