Gummy

Gummy is a state-of-the-art voice translation system developed by Tongyi Lab and unveiled at the 2024 Yunqi Conference. Unlike traditional cascadeD systems that first transcribe speech to text and then translate, Gummy uses an end-to-end neural architecture to directly map source-language speech into target-language speech or text — significantly reducing latency while preserving Fluency and contextual accuracy. It is designed for seamless integration into live communication scenariOS where speed, multilingual flexibility, and domAIn-specific precision matter most.

What it does

Gummy perfORMs real-time spoken language translation without requiring intermediate text representation. It accepts streaming audio input — such as from a microphone or conference call — and delivers low-latency translations in near real time. The model supports simultaneous reCognition and translation across diverse language pairs, including Chinese, English, Cantonese, Japanese, Korean, French, German, Russian, Italian, and Spanish. OuTPUt formats may include translated text, synthesized speech, or both, depending on deployment configuration.

Key capabilities

• Ultra-Low Latency: Translation delay is under 500 milliSEConds, enabling natural turn-taking in conversations.
• Language-agnostic mixing: Automatically detects and translates mixed-language input without requiring manual language tagging.
• Domain adaptation: Supports terminology intervention and domain Prompting to improve accuracy in specialized contexts like healthcare, legal, or technical training.
• streaming-first design: Built around a 'wait & predict' mechanism that dynamically balances translation quality and responsiveness.
• End-to-end optimization: Trained jointly on speech-to-speech and speech-to-text objectives, avoiding Error Propagation common in pipeline-based systems.

Who it is for

Gummy serves Professionals and organizations needing high-fIDElity, responsive voice translation: international conference organizers, language educators, travel tech platforms, customer support teams operating across borders, and telehealth providers delivering multilingual care. Its Architecture also makes it suitable for developers building Embedded translation features into APPs, hardware devices, or Enterprise communication tools.

How to get started

Gummy is accessible via Tongyi Lab’s official platform at tongyi.aliyun.com. Developers can integrate it using RESTful APIs or SDKs; documentation, Sample code, and sandbox environments are available for Testing. No local installation is required — all inference runs on Tongyi’s Cloud Infrastructure. Users should verify supported languages and output modes for their specific use case, as availability may vary by region and deployment tier.

Pricing and access caveat

Gummy offers free-tier access for evaluation and liMITed usage, with trial credits provided upon registration. Commercial deployments may require subscription plans or pay-as-you-go credit packages. While the core service is Freely accessible without VPN in most regions, certain advanced features or high-volume throughput may be subject to regional restrictions or authentication requirements. Pricing details, veRSIon updates, and SLA terms are not publicly Standardized and may vary based on enterprise agreements or evolving platform policies.