June 2026 Summaries

Gradium has introduced two innovative models, stt-translate and s2s-translate, to enhance real-time speech translation by combining transcription and translation into a seamless process. The stt-translate model efficiently converts spoken language into translated text across five languages without the need for intermediate transcription, while s2s-translate extends this capability to deliver end-to-end Speech-To-Speech translation by directly transforming spoken audio in one language to another using a Text-To-Speech (TTS) model. These models aim to surpass traditional three-stage translation processes by offering superior accuracy and reduced latency, as evidenced by benchmarking against gpt-realtime-translate and gemini-3.5-live-translate with higher BLEU and MetricX scores. Additionally, s2s-translate offers customizable voice output, allowing users to select or clone voices for the translated speech, thereby maintaining speaker identity and facilitating applications like dubbing and localization. This streamlined two-model architecture not only minimizes processing time but also enhances user control over the translation process, offering a significant improvement over existing solutions.

Gradium TTS has introduced a new model that significantly enhances the accuracy and naturalness of its text-to-speech output, focusing on reliably handling complex pronunciation cases, such as acronyms, numbers, and special characters, which are crucial for production voice agents. This update addresses real-world challenges encountered in live environments, such as phone-based interactions, by training on user feedback rather than standard benchmarks, resulting in improvements across multiple languages. Objective metrics and human preference tests demonstrate that the new model outperforms its predecessor and competitors like ElevenLabs in critical areas, with notable success in English and French, particularly for email addresses and time expressions. The model is designed for seamless integration, requiring no additional setup for current users, and Gradium continues to refine its capabilities by incorporating user-submitted complex cases into its development process.

In the realm of voice agents, the distinction between acoustic and semantic Voice Activity Detection (VAD) plays a crucial role in enhancing communication effectiveness and user experience. Acoustic VAD relies on signal properties like energy and spectral shape to detect speech, but can often misinterpret pauses, leading to interruptions or sluggish responses. Semantic VAD, however, incorporates language context to determine if a speaker's utterance is complete by analyzing lexical content and syntactic completeness. Gradium's Speech-to-Text (STT) system innovates by integrating turn-completion predictions directly into its audio model, allowing agents to adaptively choose response timing based on inactivity probability forecasts. This approach minimizes latency, ensures accuracy, and allows for customizable configurations to suit different conversational environments, thereby addressing common challenges in voice interaction such as early interruptions, response delays, and issues with telephony and noisy channels.