Advances in Speech Recognition: Concepts, Deep Learning Methods, and Didi’s Applications

With the rapid development of AI, intelligent voice interaction is being deployed at scale by major companies. Didi, a leading mobile‑internet platform, is actively applying speech recognition technologies—including voice recognition, dialogue understanding, and speech synthesis—to build driver assistants and smart customer‑service systems.

The core of speech recognition is converting audio signals into textual sequences. The process can be described as a search for the most probable word sequence given acoustic features X, which decomposes into a language model P(W) and an acoustic model p(X|W). Typical ASR pipelines consist of three components: acoustic model, language model, and decoder.

Classic methods rely on statistical language models such as N‑gram (based on the Markov assumption) and acoustic models like GMM‑HMM. Before deep learning, GMM‑HMM was the standard acoustic model, where GMM models the distribution of acoustic features and HMM captures temporal state transitions.

Deep‑learning breakthroughs introduced DNN‑HMM, replacing the GMM with a deep neural network while keeping the HMM decoder. End‑to‑end models further simplified the pipeline: CTC removes the need for frame‑level alignment, and attention‑based models (e.g., Listen‑Attend‑Spell) directly map audio features to text using encoder‑decoder architectures.

Transformer‑based ASR models, often pre‑trained with BERT‑style objectives (e.g., MPC), have shown improved error rates on benchmarks such as HKUST. These models handle the mismatch between the high‑frequency audio frames and the lower‑frequency textual tokens.

Despite end‑to‑end advances, practical ASR systems still require robust signal‑processing front‑ends, including echo cancellation, dereverberation, beamforming, noise suppression, and automatic gain control, to ensure reliability in far‑field and noisy environments.

Multimodal approaches combine speech and text encoders (e.g., BiLSTM‑based) with attention‑driven fusion networks to enhance tasks like emotion recognition.

In a Q&A session, deployment challenges were discussed: streaming ASR demands sub‑500 ms latency and is handled by configuring CPU concurrency, while offline batch processing can tolerate higher latency and is easier to schedule on idle machines.