Qwen3.8-LiveTranslate model release

TL;DR

Qwen released Qwen3.8‑LiveTranslate, a simultaneous interpretation system that cuts average lagging (LAAL) from 2.8 s to 2.3 s, introduces real‑time speaker separation with voice cloning, synchronized bilingual display, and long‑context disambiguation for 60 input languages.

Key Technical Advances

Interleave Architecture Improves Quality and Latency

"audio and text woven together, higher quality and lower latency. This generation recasts simultaneous interpretation as a single audio‑text interleaved stream; both the audio already heard and the translation already produced can be cached and reused." The interleaved stream lets the model treat incoming audio and generated translation as a causal sequence, enabling reuse of previously processed segments and reducing average lagging (LAAL) from 2.8 s to 2.3 s.

Real‑Time Speaker Separation and Voice Cloning

"When several people speak in turn, the model distinguishes different speakers and what each of them says, and helps the translated speech preserve each speaker’s timbre more accurately and stably." Speaker IDs are emitted alongside each translated sentence, and the Talker module synthesizes speech that mirrors the original speaker’s timbre.

Synchronized Source‑and‑Translation Output

"Bilingual alignment during interpretation supports both instant comprehension and source checking, laying a foundation for follow‑on features such as subtitle display, content organization, and retrieval." The service returns the source‑language transcript and the translation in a single response, enabling side‑by‑side bilingual screens.

Long‑Context Disambiguation

"By drawing on prior text and historical context, the model eases the ambiguity of proper nouns, references, and the like in complex scenarios, keeping expression consistent." The Thinker module incorporates earlier turns into the causal sequence, improving handling of names, terminology, and pronouns.

Model Architecture

  • Hybrid‑Mixture‑of‑Experts (MoE) Thinker–Talker design
    • Thinker: ingests video, audio, source text, and visual frames, arranging them into a single causal sequence.
    • Talker: takes the translation text and source audio to synthesize speech that preserves the original speaker’s timbre.
  • End‑to‑end streaming pipeline: audio → interleaved representation → translation → speech synthesis.

Performance Evaluation

Multi‑Speaker Long‑Audio Benchmark (Omnilingua‑MSpeaker)

  • Covers 14 language directions with multi‑speaker, long‑duration audio.
  • Qwen3.8‑LiveTranslate outperforms mainstream real‑time interpretation systems on translation faithfulness, fluency, conciseness, and Diarization Error Rate (DER).

Multilingual Real‑Time Interpretation (FLEURS audio test set)

  • Evaluated on 70 language directions.
  • Leads previous Qwen generation and competing systems on translation quality, average lagging, speech recognition accuracy, and speech synthesis quality.

No specific numeric scores were disclosed in the source; the claim is limited to relative superiority.

Supported Languages

Input audio & output text (60) Output audio (29)
Afrikaans, Arabic, Asturian, Azerbaijani, Belarusian, Bengali, Bosnian, Bulgarian, Cantonese, Catalan, Cebuano, Chinese, Croatian, Czech, Danish, Dutch, English, Estonian, Filipino, Finnish, French, Galician, Gujarati, German, Greek, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Italian, Japanese, Javanese, Kannada, Kazakh, Korean, Kyrgyz, Latvian, Macedonian, Malay, Malayalam, Marathi, Norwegian, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Slovak, Slovenian, Spanish, Swahili, Swedish, Tajik, Thai, Turkish, Ukrainian, Urdu, Vietnamese Chinese, English, German, Italian, Portuguese, Spanish, Japanese, Korean, French, Russian, Thai, Indonesian, Arabic, Vietnamese, Turkish, Finnish, Polish, Hindi, Dutch, Czech, Urdu, Filipino, Swedish, Danish, Hebrew, Icelandic, Malay, Norwegian, Persian

Using the Model via DashScope API

The blog provides a complete Python client that:

  1. Opens a WebSocket to wss://{workspace_id}.cn-beijing.maas.aliyuncs.com/api-ws/v1/realtime?model=qwen3.8-livetranslate-flash-realtime.
  2. Configures the session to request text and optional audio output, enabling speaker IDs and source‑language ASR.
  3. Streams microphone PCM chunks (16 kHz, 16‑bit) to the server.
  4. Receives incremental translation text, synthesized audio, and speaker identifiers.
  5. Plays back synthesized audio in a background thread (when audio_enabled is true).

Key code excerpts (excerpted for brevity):

config = {
    "session": {
        "output_modalities": ["text", "audio"] if self.audio_enabled else ["text"],
        "input_audio_format": "pcm",
        "output_audio_format": "pcm",
        "translation": {"language": self.target_language}
    }
}
await self.ws.send(json.dumps(config))

The client handles events such as response.audio.delta, response.text.delta, and session.finished to deliver real‑time translation and clean shutdown.

Demonstrations

The blog showcases two dialogues from Journey to the West and a set of homophone examples, illustrating:

  • Accurate speaker attribution and stable voice cloning across turns.
  • Synchronized bilingual screen output.
  • Context‑aware disambiguation of ambiguous terms using visual cues.

Future Directions

Qwen outlines three research priorities:

  1. Latency compression – push end‑to‑end latency toward the theoretical limit of simultaneous interpretation.
  2. Long‑term memory across sessions – retain conversational context across multiple sessions for the same project and participants.
  3. Expanded language coverage – add more long‑tail languages and regional dialects to achieve universal real‑time interpretation.

Citation

If you use Qwen3.8‑LiveTranslate in research or products, cite the blog entry as follows:

@misc{qwen38livetranslateblog,
  title = {Qwen3.8-LiveTranslate: Names the speaker. Carries the meaning.},
  url = {https://qwen.ai/blog?id=qwen3.8-livetranslate},
  author = {Qwen Team},
  month = {September},
  year = {2026}
}

All statements are drawn directly from the Qwen blog post dated September 18 2026.

Sources