Intelligent Voice Robot Architecture, Core Technologies, and Enterprise Applications

Intelligent voice robots combine speech recognition, semantic understanding, and speech synthesis to enable multi‑turn human‑machine conversations, and are widely used in marketing, product promotion, and service notifications. The speaker, Li Zhong, an algorithm architect at 58.com, shares the system architecture and deep dives into two core modules: intent recognition and dialogue management.

The overall voice interaction flow distinguishes between two input types: raw speech signals and event/command signals (e.g., long silence, hardware actions). For speech input, the pipeline includes voice type classification, ASR, NLU (regular‑expression matching, intent classification, slot filling), and a dialogue manager that selects either script‑driven or knowledge‑base‑driven responses, finally generating TTS or pre‑recorded audio.

Event or command inputs trigger predefined response strategies in the dialogue manager, such as handling DTMF key presses or long‑silence reminders.

Voice preprocessing consists of two modules: (1) a "clear‑speech" binary classifier that uses 10‑second audio segments, Fbank features, and a simplified VGG network to filter out unintelligible speech, improving downstream intent accuracy by ~29%; (2) Voice Activity Detection (VAD), evolving from double‑gate energy/zero‑crossing methods to Google WebRTC VAD and now a deep‑learning‑based VADNet, which raises frame‑level detection accuracy by 46%.

Intent recognition employs a TextCNN model (compared with fastText and LSTM) to classify 19 intent categories, with an optional BERT‑based encoder that adds about 2% accuracy. Standard question matching uses DSSM and BERT similarity models to retrieve answers from a knowledge base when user queries fall outside scripted flows.

Slot extraction is performed with an IDCNN+CRF model to identify entities such as time, location, or product type, enabling precise dialogue branching.

In 58.com, the voice robot is deployed primarily as a telephony robot, handling outbound calls, multi‑turn dialogues, and intent detection across scenarios like notifications, follow‑ups, verification, sales, and opportunity mining, delivering cost‑effective, 24/7 service.

Several real‑world cases illustrate business impact: (1) sales efficiency – filtering unqualified leads raises qualified prospect ratio by 46×; (2) opportunity mining – proactive calls recover lost sales and generate platform fees; (3) customer‑service automation – voice bots replace manual reminders for member verification; (4) operational efficiency – B‑side merchants receive timely prompts to improve response rates; (5) campus recruitment – automated phone notifications and key‑press analysis streamline interview scheduling; (6) AI‑driven interviews – bots conduct preliminary video interviews, feeding results back to recruiters.