Uploaded April 2026 | Updated November 2026, 3 days ago
What does it take to run real‑time speech recognition for competitive mobile games—without relying on the cloud?
In this GDC session, Tencent GVoice and Qualcomm share how they brought on‑device AI ASR into Honor of Kings, one of the world’s most played mobile games, to deliver faster, more reliable in‑game voice experiences under real‑world conditions.
Rather than presenting a generic ASR solution, this talk walks through one concrete production case—from noisy public environments to regional dialects—showing how cloud‑scale speech pipelines were redesigned for mobile, latency‑critical gameplay.
You’ll learn:
- Why latency and network round‑trips break voice in competitive mobile games
- How Tencent distilled and compressed large ASR models for on‑device use
- Techniques for domain‑specific speech training (game jargon, slang, dialects)
- Using Snapdragon NPU, CPU, and HVX to balance accuracy, memory, and power
- Real performance results from Honor of Kings production deployment
This session is especially relevant for teams building voice‑driven gameplay, accessibility features, or AI‑powered moderation on mobile.
Speakers:
Lily Zhu (Tencent)
Jane Tian (Qualcomm Technologies)
Key sections:
00:00 Why voice is mission‑critical in mobile MOBAs
03:02 GVoice background and Honor of Kings requirements
07:37 From cloud ASR to on‑device inference
09:44 In‑domain training for game‑specific speech
12:55 Distillation and streaming model design
15:39 Quantisation and NPU‑friendly deployment
18:45 Performance, memory, and latency results
19:49 Qualcomm AI stack for on‑device ASR
21:56 Developer resources and next steps
Subscribe to get news and updates on Snapdragon's game developer solutions: qualcomm.com/support/contact/forms/snapdragon-gaming
What does it take to run real‑time speech recognition for competitive mobile games—without relying on the cloud?
In this GDC session, Tencent GVoice and Qualcomm share how they brought on‑device AI ASR into Honor of Kings, one of the world’s most played mobile games, to deliver faster, more reliable in‑game voice experiences under real‑world conditions.
Rather than presenting a generic ASR solution, this talk walks through one concrete production case—from noisy public environments to regional dialects—showing how cloud‑scale speech pipelines were redesigned for mobile, latency‑critical gameplay.
You’ll learn:
- Why latency and network round‑trips break voice in competitive mobile games
- How Tencent distilled and compressed large ASR models for on‑device use
- Techniques for domain‑specific speech training (game jargon, slang, dialects)
- Using Snapdragon NPU, CPU, and HVX to balance accuracy, memory, and power
- Real performance results from Honor of Kings production deployment
This session is especially relevant for teams building voice‑driven gameplay, accessibility features, or AI‑powered moderation on mobile.
Speakers:
Lily Zhu (Tencent)
Jane Tian (Qualcomm Technologies)
Key sections:
00:00 Why voice is mission‑critical in mobile MOBAs
03:02 GVoice background and Honor of Kings requirements
07:37 From cloud ASR to on‑device inference
09:44 In‑domain training for game‑specific speech
12:55 Distillation and streaming model design
15:39 Quantisation and NPU‑friendly deployment
18:45 Performance, memory, and latency results
19:49 Qualcomm AI stack for on‑device ASR
21:56 Developer resources and next steps
Subscribe to get news and updates on Snapdragon's game developer solutions: qualcomm.com/support/contact/forms/snapdragon-gaming










