Manta: a tool developed by Stingray to reduce translation costs
By Stingray | Published on December 11, 2024
Problem and objectives
Problem
- Faced with the challenges of translating diverse and multilingual content, we realized the need to develop an internal tool. Reliance on outsourcing resulted in high costs and extended lead times, hampering our efficiency. To guarantee impeccable quality in various areas, we designed Manta, a tailor-made solution to optimize our resources and improve our responsiveness.
- High costs: Outsourcing translation services results in significant expenses.
- Extended deadlines: The numerous return trips lengthen processing times.
- Multilingual complexity: Managing multiple languages requires a rigorous approach.
- High volume: Managing subtitles and other content requires special attention to maintain quality.
- Quality requirements: Legal documents and concert descriptions require impeccable precision.
Objectives
- Cost reduction: Reduce expenses by eliminating outsourcing.
- Save time: Speed up the translation process for greater efficiency.
- Internalization of operations: Bring translations in-house for better control.
- Quality assurance: Guarantee consistent quality for critical documents.
- Fast delivery: Accelerate the delivery of translations to meet business needs.
- Extensive multilingualism: Cover a wide range of languages for diverse content.
Proposed solution
Solution
We created Manta, an advanced AI translation agent that goes beyond simple translation. It offers tailored language profiles and follows specific guidelines for each type of content, incorporating style and terms specific to each domain. Manta translates faster than traditional methods, while ensuring high quality and contextual accuracy.
Principle
Manta uses AI algorithms to analyze and translate texts taking into account linguistic and cultural context. The system allows users to choose the appropriate language profile and applies specific guidelines depending on the type of content. It also benefits from improvements brought by new AI models.
Manta takes care to maintain the original synchronization of the subtitles while respecting the specificities of the dialogues, guaranteeing a natural translation that is faithful to the context. Additionally, to preserve their artistic integrity, song lyrics and titles, such as those by Radiohead, are not translated.
User experience
Manta offers an intuitive interface, allowing users to choose document type, language, profile and lexical style. Accessible to all employees, it facilitates the creation of precise profiles.
Role of AI
AI has been crucial to overcoming our translation challenges. With Manta, we were able to:
- Reduce costs: Significant reduction in subcontracting expenses.
- Accelerate deadlines: Increased translation speed, reducing delays.
- Maintain high quality: Advanced algorithms for consistent quality.
- Internalize the process: Better control and smooth integration of translations.
- Deliver in multiple languages: Effective management of linguistic diversity.
Impacts
Main outcomes
Enhanced Trust in AI: Manta's success, increasing trust in AI and adoption of new technologies.
Resource Optimization: Savings to redirect resources to other projects.
Automation Incentive: Success motivating increased integration of AI for other processes.
Challenges takled
Resistance to change and concerns about the quality of AI translations have posed challenges. We optimized the prompts to improve quality. An internal blind test revealed a preference for AI translations, with few errors. A customer test provided valuable feedback to adjust our prompts. Internal testing prior to delivery was crucial to ensuring quality, enabling us to achieve our initial goals.
Conditions for success
Team mobilised
The project started as a quick proof of concept, with a developer dedicated to improvements. A/B testing and blind surveys assessed quality. After validation, the translations were sent to clients. This approach has led to gradual adoption. We continue to improve the system with GPT-4o, providing better quality and lower costs, while using the language model to self-critique.
Collaborations
Customers have provided valuable feedback to improve our prompts. Employees actively sought to integrate the tool and suggest improvements, ensuring successful adoption.
Funding received
No funding program was used for this project.
Project stages
Design and development: Proof of concept with a dedicated developer.
Internal testing: A/B testing and blind surveys.
Customer feedback: Customer feedback to adjust prompts.
Gradual adoption: Managing resistance and encouraging adoption.
Continuous improvements: Integration of new technologies like GPT-4o.
Project stages - to come
We focus on continuous improvement, with the aim of increasing the available profiles and integrating new language models to improve the quality and efficiency of translations.