# Qwen3.8-LiveTranslate Tested: Real-Time AI Interpretation, With Glitches > Source: > Published: 2026-09-21 00:00:00+00:00 # Qwen3.8-LiveTranslate Tested: Real-Time AI Interpretation, With Glitches A hands-on test of Qwen3.8-LiveTranslate across a dozen languages shows solid translation quality but real turn-taking and latency problems. ## What is Qwen3.8-LiveTranslate? Qwen3.8-LiveTranslate is a real-time simultaneous interpretation model from the Qwen team, built to listen to speech in one language and produce translated speech and text in another, live, while the speaker is still talking. It supports input and output across 60 languages, with 29 of those getting full voice output and the rest limited to text. That’s a wide net for a single model, covering most of the world’s major language families in one system. It’s currently available only through an API, with no confirmed timeline for an open-source release. ## TL;DR - **Qwen3.8-LiveTranslate** is a real-time speech-to-speech and speech-to-text translation model supporting 60 languages, with 29 offering full voice output. - A hands-on test using the official API and script showed that **short, single sentences translate cleanly** , but the model struggles once speech continues past that. - The biggest weakness observed was **turn-taking** : the system had trouble detecting when the speaker had finished talking, cutting off mid-translation or stalling. - **Interruption handling didn’t work** in testing either, meaning users can’t easily correct or redirect the model mid-stream the way they could in a live conversation. - Under the hood, the model uses a **hybrid mixture-of-experts architecture** with a “thinker” component that streams source and translated text together, and a “talker” component that generates speech in the original speaker’s cloned voice. - Qwen claims the model **matches or beats GPT Realtime Translate and Gemini 3.5 Live Translate** on faithfulness, fluency, and consistency benchmarks, and that its diarization error rate is roughly half of GPT’s and a third of Gemini’s. - Translation quality across tested languages (Russian, Arabic, Dutch, German, Urdu, Hindi, Indonesian, Spanish, Portuguese, Persian) was generally accurate for short phrases, but the **real-time promise doesn’t hold up out of the box** . ## Remy doesn't write the code. It manages the agents who do. Remy runs the project. The specialists do the work. You work with the PM, not the implementers. ## How does Qwen3.8-LiveTranslate actually work? The architecture behind Qwen3.8-LiveTranslate is a hybrid mixture-of-experts model split into two functional pieces. The first, called the “thinker,” processes incoming speech and streams the source transcription and its translation together rather than waiting for a full sentence to finish. The second, the “talker,” takes that output and generates spoken audio, aiming to preserve the original speaker’s vocal characteristics rather than defaulting to a generic voice. Instead of generating text first and then audio afterward, the system interleaves both streams, producing text and speech close to simultaneously. Qwen highlights three specific capabilities meant to set this apart from a standard translation pipeline: - **Speaker separation and voice cloning** , so each speaker in a multi-person conversation keeps a distinct, cloned voice in the output. - **Synchronized side-by-side source and translation** , showing both the original and translated text together instead of only the final result. - **Context-aware disambiguation** , using prior conversation context to resolve references like a nickname pointing back to a full name mentioned earlier. On paper, this is a genuinely ambitious design. It’s trying to solve translation, diarization, voice preservation, and low-latency streaming all at once, in a single model. ## Does it work in real-time, out of the box? In practice, not consistently. A hands-on test ran the model through its official API using Qwen’s own sample script, feeding it live microphone input and picking different target languages for each pass. Short sentences, things like a single line of courtroom dialogue, came through fine. But as soon as the input extended into anything resembling a real, continuous speech pattern, the system started breaking down. The core issue was turn-taking: the model had difficulty recognizing when the speaker had actually finished a thought and it was time to deliver or continue the translation. In several tests, the output simply stopped partway through, or lagged noticeably behind the live input before catching up late. Interruption handling, the ability to jump in and redirect or correct the model mid-stream, also didn’t function as expected during testing. This matters because the entire pitch of a “live” interpretation model rests on it behaving like a human interpreter: continuous, responsive, able to handle natural pauses and overlapping speech. Based on this testing, achieving that kind of fluid real-time behavior would require building a custom pipeline around the model rather than relying on the default implementation, at least at this stage. ## How good is the translation quality across languages? Quality, judged purely on the text/speech that did come through, was generally solid for short inputs. The same test sentence, a courtroom line about a defendant being present, was run across Russian, Arabic, Dutch, German, Urdu, Hindi, Indonesian, Spanish, Portuguese, and Persian. Most outputs were rated as accurate or close to it, with Indonesian, Spanish, Portuguese, and Persian singled out as particularly clean. Urdu and Hindi were called “okay” but not flawless. ## Other agents ship a demo. Remy ships an app. Real backend. Real database. Real auth. Real plumbing. Remy has it all. The caveat is that all of these were single, short sentences. Longer or more complex phrasing wasn’t where the model succeeded in this test, it was where the turn-taking problems surfaced. So the underlying translation engine appears capable, but the system wrapped around it (deciding when to start, stop, and hand off) is the weaker link right now. ## Is Qwen3.8-LiveTranslate worth using today? For anyone needing a fast, accurate translation of short, discrete utterances (a single question, a short command, a brief exchange) it performs reasonably well across a broad set of languages. For anyone hoping to drop it into a live conversation, meeting, or interpretation scenario expecting smooth, continuous, human-interpreter-like behavior, it currently falls short without additional engineering work on top of the default setup. That gap between the marketing framing (real-time, simultaneous, naming the speaker, carrying the meaning) and the out-of-the-box experience is the main thing to weigh. The model’s ambitions, and the technical approach behind it, are legitimate advances for the field. But “real-time” is doing a lot of work in that name, and based on hands-on testing, it isn’t fully earned yet without custom pipeline work around voice activity detection, turn segmentation, and interruption handling. ## How does it compare to GPT and Gemini’s live translation? Qwen has published benchmark claims stating that Qwen3.8-LiveTranslate matches or beats GPT’s real-time translation and Gemini’s 3.5 Live Translate model on three axes: faithfulness (how accurately meaning is preserved), fluency (how natural the output sounds), and consistency across a conversation. Qwen also claims its diarization error rate, meaning how well it distinguishes between different speakers, is roughly half of GPT’s and about a third of Gemini’s. Those are Qwen’s own reported figures, and independent, controlled benchmark comparisons weren’t part of the hands-on testing here. They should be read as the vendor’s claims rather than a verified third-party result. What the hands-on test does confirm is that this is a fast-moving, competitive corner of AI: multiple major labs are now racing to build real-time speech translation with speaker preservation, and the field is still clearly in an early, unsettled stage. ## Frequently Asked Questions ### What languages does Qwen3.8-LiveTranslate support? It supports input and output across 60 languages total. Of those, 29 languages get full voice output, meaning the model speaks the translation aloud in a cloned version of the original speaker’s voice. The remaining languages are supported as text-only output. ### Is Qwen3.8-LiveTranslate open source? No, as of testing it’s available only through Qwen’s API. There’s no confirmed timeline for an open-weight or open-source release. ### Does Qwen3.8-LiveTranslate clone the speaker’s voice? Yes, that’s one of its core design goals. The “talker” component of its architecture is built to preserve the original speaker’s vocal characteristics in the translated audio output, and the model also separates and clones voices individually when multiple speakers are present. ### Why does the translation stop working on longer speech? Based on hands-on testing, the issue isn’t translation accuracy itself but turn-taking: the model has trouble reliably detecting when a speaker has finished talking. On short, single sentences it performs well, but continuous or longer speech exposed lag, stalling, and failed interruption handling. ### How does Qwen3.8-LiveTranslate’s architecture work? It uses a hybrid mixture-of-experts design split into two parts: a “thinker” that streams the source transcription and its translation together, and a “talker” that converts that into spoken audio while preserving the original speaker’s voice. Text and audio are generated in an interleaved stream rather than sequentially.