Baidu Speech Synthesis: Balancing Trade‑offs and Opening the Platform to Developers

At the 52nd Baidu Technology Salon on July 26, Li Xiulin, who leads Baidu’s speech synthesis system development, explained why speech synthesis matters: it fulfills the long‑standing human dream of making machines speak and, more importantly, provides a natural voice interface for special groups who face barriers to accessing modern information.

Speech synthesis, also known as text‑to‑speech (TTS), underpins many everyday products such as car navigation, smartphone assistants, e‑reading apps, and emerging wearable devices. Baidu has been researching TTS since early 2013, launched its open platform in April 2014, and integrated the technology into the Baidu search box by July 2014.

Li highlighted the technical challenges of Chinese TTS: over 1,400 tone‑marked syllables generate millions of contextual variations. Baidu addressed this by applying phonetic and linguistic knowledge to classify contexts and compress the feature space. For example, consonants (initials) are grouped from dozens of types down to a handful based on articulation method and place.

The system uses initials and finals as basic units, reducing the unit inventory and further shrinking the feature space. Trade‑offs are made among recording corpus size, speaker diversity, model training sufficiency, scalability, and the balance between subjective perception and acoustic parameters, providing greater flexibility for system optimization.

In April 2014 Baidu opened the speech synthesis service to developers for free via its Speech Open Platform. Developers can download the SDK, call the provided APIs, and let Baidu’s online service handle data, machines, and network management, allowing them to focus on application logic.

Since the public launch in October 2013, the platform has attracted major mobile apps such as Momo, Qunar, and Air China. The platform offers a complete solution that enables developers to integrate advanced synthesis and recognition technologies at low cost.

Technically, Baidu’s front‑end leverages massive corpora for natural language understanding, including intelligent word segmentation, high‑precision polyphone handling, and accurate prosody prediction. A refined acoustic model built with an HMM framework and extensive optimization yields a robust yet expressive system. Fast unit pre‑selection and multi‑level cost optimization select the best units for concatenation.

Li concluded that Baidu’s mission is to make information more equally and conveniently accessible, and that speech synthesis is a key means to achieve this goal.