Gradium Launches stt-translate and s2s-translate

stt-translate and s2s-translate debut with new architecture

Gradium launched Stt-translate and s2s-translate, the latest real-time speech translation models designed to tackle the latency and accuracy bottlenecks that often plague standard speech-to-speech pipelines. They're fast. But the company isn't sacrificing quality, collapsing a traditional three-step process into a two-step flow to target competitive performance against major industry players.

The models support five languages: English, French, German, Spanish, and Portuguese. This selection enables 20 different language pairs. It allows for flexible translation across these specific tongues. Instead of relying on a multi-stage handoff where speech is transcribed, then translated, and finally synthesized, the new system performs transcription and translation in a single pass, and it's designed to remove the delays inherent in managing multiple intermediate steps.

Engineering a faster translation pipeline

Traditional speech-to-speech stacks require three distinct stages. It's complicated. First, a model converts speech to text, then a separate engine handles the text-to-text translation, and finally a third system generates the output speech. But each handoff between these stages acts as a point of friction, so they're increasing total latency.

The new approach eliminates the dedicated text-to-text stage entirely. It's simpler now. By integrating transcription and translation into one process, the system reduces the number of moving parts, and according to information provided by Gradium, this architecture relies on the Hibiki-Zero framework. So the models use Reinforcement Learning to balance the tradeoff between response speed and output accuracy.

Performance metrics and real-world results

The developers used two benchmarks: BLEU and MetricX. BLEU focuses on lexical fidelity by measuring n-gram overlap with human translations, but MetricX acts as a neural quality metric that mimics human evaluation instead. So it's a key distinction. The results show the following.

• The average latency for the s2s-translate pipeline is 3.0 seconds.
• This speed outperforms gpt-realtime-translate, which sits at 3.6 seconds.
• Gemini-3.5-live-translate remains slightly faster at 2.9 seconds.
• Gradium reports lead positions over gemini-3.5-live-translate on both BLEU and MetricX benchmarks.

While the numbers suggest a lead in accuracy, the company notes that these results come from a proprietary dataset focused on everyday conversational topics like travel and work. Real-world performance might vary depending on the complexity of the speech.

New features for voice and accessibility

Speed matters. But this system offers more , it adds output voice control. Users can pick specific voices or use cloning technology to mimic the speaker's original tone, and the whole thing runs over a single duplex WebSocket that makes integration simpler for developers. The server handles the entire pipeline. It streams both the synthesized audio and the translated transcript back to the client as they're generated.

The approach draws on the Hibiki-Zero framework, optimizing low latency and high accuracy jointly through Reinforcement Learning.

This integration removes the need for developers to wire together separate speech-to-text and text-to-speech services. So the platform accepts input audio in PCM format at 24 kHz and 16-bit signed mono. It outputs at 48 kHz. And it also includes support for other audio formats like Opus, WAV, mu-law, and A-law, which can be quite handy for various applications.

Potential use cases for live translation

The practical applications for this technology span several industries. Live dubbing is a primary target, where a keynote speaker’s voice can be cloned and translated into another language in real time. Similarly, multilingual voice agents can route support calls through the pipeline, allowing an English-speaking agent to communicate with a German caller without needing a human interpreter.

Accessibility is another core area. For scenarios where audio synthesis is not required, stt-translate can operate independently to provide live translated captions. This removes the overhead of generating audio while still providing the necessary text stream for participants in meetings or broadcasts.

Integration and future outlook

Developers can access these capabilities via a Python SDK. It abstracts the connection details into just a few lines of code. The client handles the streaming of audio buffers directly to the API, and it returns both the text and audio results, so you don't need to manage multiple connections. But it's a straightforward implementation.

The current release covers five languages. But the limit to just 20 pairs might be a problem for users who need a much wider linguistic scope for their work or personal projects. It's a short, two-model path. That decision reflects a broader industry trend of tightening the feedback loop in real-time AI, and as the technology moves into broader use, we'll likely see this simplified stack tested against far more diverse conversational datasets.