Exhibitor News

Subpage Hero

Subpage Hero


28 Jun 2023

A Guide to Building High-Quality Text-to-Speech Datasets

CodeLink Stand: N23
A Guide to Building High-Quality Text-to-Speech Datasets

In today's rapidly evolving digital landscape, text-to-speech AI has emerged as a game-changer, powering voice assistants, audiobook narrations, and GPS systems. But what sets these AI-generated voices apart, making them sound so natural and human-like?

The secret ingredient is high-quality data used to train text-to-speech AI models.

In this blog, we'll delve into the significance of data collection, examine various methods and tools for collecting and preparing data, discuss challenges and ethical considerations, and look at real-world examples.

The Crucial Role of Data in Text-to-Speech AI

Data is the lifeblood of text-to-speech AI models. The more diverse, representative, and unbiased the data used to train these models, the more accurate and natural-sounding the resulting speech.

High-quality data ensures AI-generated speech is inclusive, catering to different accents, dialects, and languages. This broadens the reach of AI voices, making them accessible to a wide range of users while minimizing errors and inaccuracies.

Harnessing Cutting-Edge Tools for Data Collection and Annotation

Efficient data collection and annotation are crucial for developing accurate and high-quality text-to-speech AI models. Pioneering tools, such as annotation software and automatic speech recognition (ASR) services, have revolutionized the process. These tools not only automate the alignment of audio and text but also offer user-friendly interfaces for manual adjustments, ensuring the precision and quality of text-to-speech models.

One particularly helpful annotation tool for text-to-speech (TTS) data simplifies the alignment of audio and text, making it easy for users to create audio-text pairs:

  1. Upload an audio file.
  2. Input the corresponding text.
  3. The system processes the audio and divides it into smaller segments.
  4. An ASR service assigns text to each segment.
  5. Users can edit the text and align the audio using a slider within the app.
  6. The resulting audio-text pair is saved in Firebase storage, retaining the original audio file's name.

By streamlining the annotation process, this tool enhances the accuracy and quality of TTS models, making them more efficient and valuable across a wide array of applications.

Creating Your Own Text-to-Speech Datasets

Annotation tool for TTS datasets: https://storage.googleapis.com/strapi-bucket-production/image1_85455d075f/image1_85455d075f.png

Ready to build your own AI-generated speech? Codelink has an open source a tool for collecting and processing speech data in this Github repository. This is a guide to collect data, set up a development environment, and execute AI text-to-speech using Firebase and Google Cloud Platform.

To run the project, you'll need to set up both the Firebase project and Firebase storage. Key tasks include creating a Firebase project, setting up Firebase storage, and configuring a Firebase Cloud Function. Additionally, you'll need a Google Cloud Platform project to operate the Text-To-Speech project. For further guidance, consult the Firebase documentation.

By following steps annotated in the Github repository, along with configuring essential files and deploying services to Cloud Run and Firebase, you can ensure seamless development and execution of AI text-to-speech projects using Firebase and Google Cloud Platform.

Innovative annotation tools play a pivotal role in building high-quality text-to-speech AI models. By automating the alignment of audio and text and offering user-friendly interfaces for manual adjustments, these tools contribute to the precision and quality of AI-generated speech.

Collaborative Efforts: Sharing Data to Drive Innovation

Collaboration and data sharing play a vital role in advancing AI-generated speech. Embracing open-source approaches and sharing data from diverse sources (i.e., audiobooks, news articles, podcasts), companies like Google and Amazon contribute to the development of AI voices while fostering innovation and inclusivity within the field.

Open-source initiatives, like Google's decision to make its text-to-speech dataset publicly accessible, enable researchers and developers worldwide to enhance their AI models. By sharing data and tools, the AI community can collectively improve models, ensuring diverse perspectives and voices are represented in AI-generated speech.

This collaborative approach fosters a vibrant and inclusive environment, thereby greatly enriching the AI community. It not only guarantees the representation of diverse perspectives and voices in AI-generated speech but also brings about remarkable advantages for the entire community.

Case Study: A Vietnamese Dataset for Text-to-Speech

While numerous high-quality voice datasets exist for English, there is a notable gap in the availability of a high-quality Vietnamese dataset.

A Vietnamese dataset for text-to-speech (TTS) has been released using the advanced annotation tools. This dataset consists of 10,000 audio-text pairs recorded by a professional voice actor and sourced from news articles.

Designed for training TTS models in the Vietnamese language, the dataset exemplifies the potential of high-quality data in creating accurate and natural-sounding AI-generated speech across different languages.

Báo Nói Application: Integrating AI Text-to-Speech into Business

Presenting the Baonoi app, your gateway to the future of Text-to-Speech (TTS) technology. Visit us at https://baonoi.ai and unlock a unique and interactive experience that leverages an exceptionally curated dataset to educate a state-of-the-art TTS model.

Countless enterprises are reaping the rewards of AI-generated speech, fueled by top-tier data. Using progressive APIs and a multitude of data streams, such as audiobooks, they craft unique AI voices that represent their brands.

By centering on data drawn from authentic scenarios, businesses are breathing life into AI-generated speech, enhancing its naturalness and precision. The TTS landscape has seen the development of numerous models, and the quality of datasets remains the primary catalyst of this innovation.

TTS technology not only enables businesses to amplify the value of their offerings but also fosters consistent evolution in AI-generated speech technology applications. It enhances user interaction, making their experiences more seamless and enjoyable. Step into the future of TTS with the Baonoi app - where innovation meets user satisfaction.

View all Exhibitor News


Platinum Sponsor


Theatre Sponsors

Gold Sponsors

Silver Sponsors

Bronze Sponsors


Data & Generative AI Learning Partner


Women In Tech Symposium In Support of

Korea Singapore Innovative Enterprise Seminar Partners

Media Partners

Community & Content Partners

Prize Sponsors

Knowledge Partner

frost & sullivan


Official Partner Hotel

Held In

Supported By

Subscribe to Our Newsletter!