Overview of Speech and Semantic Recognition Technologies Presented at the Tencent Cloud+ Community Developer Conference

On December 15, Tencent Cloud hosted the inaugural "Tencent Cloud+ Community Developer Conference" in Beijing. The event, themed "New Trends·New Technologies·New Applications," gathered more than 40 technical experts and attracted over 1,000 developers to discuss hot topics such as artificial intelligence, big data, IoT, mini‑programs, and DevOps.

Voice and semantic recognition play a pivotal role in AI today. WeChat Zhiling focuses on research and productization of speech technologies, offering AI‑driven voice recognition that supports real‑time transcription, on‑site interpretation, and other functions. This talk examined the evolution of Zhiling’s voice technology and its application principles in mobile products and various solutions.

From a technical perspective, speech recognition consists of several modules: feature extraction, acoustic modeling, lexicon, language modeling, and decoding. Feature extraction converts raw audio into suitable features (e.g., MFCC, PLP). The acoustic model maps these features to phoneme sequences. A lexicon translates phoneme sequences into words, and a language model uses contextual word relationships to form complete sentences. Decoding combines these components to produce the final transcription.

History of Speech Recognition Technology

Traditional systems (pre‑2009) relied on MFCC/PLP features and HMM‑GMM acoustic models, where HMM handled temporal dependencies and GMM performed frame‑level classification.

After 2009, deep neural networks (DNN) were introduced, dramatically improving performance. Subsequent advances incorporated CNNs, LSTMs, and other deep‑learning techniques for both acoustic and language modeling.

Around 2014, Connectionist Temporal Classification (CTC) eliminated the need for HMMs by allowing the network to model sequences directly.

Later, end‑to‑end approaches such as encoder‑decoder architectures with attention (originating from machine translation) further simplified the pipeline, merging acoustic and language modeling into a single neural network.

WeChat Zhiling Voice Platform

Founded in 2011, the Zhiling team now has 30 members and focuses on speech recognition, speech synthesis, speaker verification, and speech assessment. Their customers span both B2C apps and B2B services.

The front‑end signal‑processing pipeline includes Voice Activity Detection (VAD) to filter non‑speech segments, audio event classification to discard laughter or music, noise reduction for background sounds, and speaker diarization to isolate individual speakers.

On the back‑end side, Zhiling continuously collects and augments large‑scale speech data, leveraging a massive GPU cluster for multi‑machine, multi‑card parallel training and decoding. Acoustic models are selected per scenario, while language models incorporate online LM re‑estimation, RNN‑LM, real‑time updates, and error‑correction mechanisms.

Performance figures reported include a 97% recognition rate for near‑field scenarios, ~90% accuracy for long‑form transcription, and 87‑88% accuracy in noisy environments such as subways and buses.

Unique modeling techniques include a multilingual (Chinese‑English) mixed model to boost accuracy in code‑switching contexts, and a customizable language model that quickly adapts to domain‑specific vocabularies.

WeChat Zhiling Cloud Cases

Zhiling’s solutions are deployed in over 50 mobile apps, handling roughly 400 million daily requests. They also power telephone‑based services for verticals such as transportation, finance, education, and insurance, with a daily cloud usage of about 30 k hours and options for private‑cloud deployment.

Tencent’s simultaneous‑translation service (Tencent Tongchuan) provides bilingual subtitles and meeting minutes for international conferences, having served events like the Bo’ao Forum, World AI Conference, and the first China International Import Expo.

In the legal sector, Zhiling offers speaker role identification and microphone‑array processing for courtroom interrogations and police inquiries.

Finally, Tencent Cloud Xiaowei integrates Zhiling’s voice interaction capabilities into smart hardware such as speakers, cars, robots, and TVs.