v1.4 / 4B v1.3: Japanese and more languages

README.md

@@ -11,6 +11,8 @@ pinned: false
 
 **A System One model for speech.** Jev made decision models the idea of the year: typed, probabilistic answers instead of generated text. DuplexJev brings that to raw audio, where a text decision model can't go because the transcript drops the tone. It answers the small closed questions a voice agent needs about every user turn (*has the user finished? should I reply, keep listening or just say "mm-hm"? is this a barge-in? emotion, gender, intent*) as a single token from a frozen LLM that hears the audio directly. **10 decisions in 92 ms**, 21× faster than ASR → LLM, under $0.50 per million decisions (4B).
 
+🆕 **Oct 22 · DuplexJev now hears Japanese.** 32B-Para v1.4: Japanese error 32.7 → 11.6% (FLEURS), conversational Japanese 38.2 → 11.1%; Korean 31.8 → 7.6%, Vietnamese 41.7 → 14.6%; Mandarin conversational laughter 89, breath 92, stressed word 85 on held-out speakers. 4B-Para v1.3 brings Japanese to 17.2%. [Release note](https://github.com/adventists-ai/duplexjev/blob/main/research/2026-10-japanese.md)
+
 🏆 **State of the art in paralinguistic understanding on public benchmarks:** DuplexJev-32B-Para v1.3 ranks first with **50.2** on LISTEN (EACL 2026), StressTest and MSPB combined (1,035 items, none from our training corpora), ahead of Gemini 3.8 Flash 47.8, Qwen3-Omni 45.0, Gemini 3.1 Pro 41.0, GPT-audio-1.5 40.9, Gemini 3.8 Live 40.0 and GPT-realtime-2.1 35.6; Gemini 3.8 Flash stays ahead on StressTest alone (89.0 vs 86.5). Details and every prediction: [benchmarks/paraling_compare](https://github.com/adventists-ai/duplexjev/tree/main/benchmarks/paraling_compare).
 
 🏆 **Turn-taking on par with dedicated detectors:** DuplexJev-32B-Para scores 95.5 on the Easy-Turn test set
@@ -21,10 +23,10 @@ while answering any other question in the same pass. 📄 Paper: [arXiv:2610.026
 
 | model | for | built on | qa100 | Easy-Turn | gender | emotion | licence |
 |---|---|---|---:|---:|---:|---:|---|
-| ⭐ [**DuplexJev-32B-Para**](https://huggingface.co/adventists-ai/DuplexJev-32B-Para) 🆕 v1.3 | best overall and best live turn-taking; widest set of abilities: much better stress (StressTest 86.5), accents, laughter, breathing, coughs, sighs, speaking style, same speaker (one 80 GB GPU) | DuplexJev-32B-Turn + para LoRA | 96 | 94.4\* | **93.8** | **92.1** | CC BY-NC 4.0 |
-| ⭐ [**DuplexJev-4B-Para**](https://huggingface.co/adventists-ai/DuplexJev-4B-Para) 🆕 v1.1 | the same on ~10 GB, and also holds a normal conversation; `duplexjev` default | DuplexJev-4B-Turn + para LoRA + conversation replay | 77 | 91.0\* | 89.2 | 90.0 | CC BY-NC 4.0 |
+| ⭐ [**DuplexJev-32B-Para**](https://huggingface.co/adventists-ai/DuplexJev-32B-Para) 🆕 v1.4 | best overall, widest set of abilities: stress (StressTest 86.5 in v1.3), accents, laughter, breathing, coughs, sighs, speaking style, same speaker; 🆕 **Japanese** (FLEURS error 32.7 → 11.6%), Korean, Vietnamese, Thai, Arabic, Russian, and Mandarin conversational paralinguistics (one 80 GB GPU) | DuplexJev-32B-Turn + para LoRA + language LoRAs | **97** | 94.1\* | 92.1 | 89.2 | CC BY-NC 4.0 |
+| ⭐ [**DuplexJev-4B-Para**](https://huggingface.co/adventists-ai/DuplexJev-4B-Para) 🆕 v1.3 | the same on ~10 GB, and also holds a normal conversation; 🆕 **Japanese** (FLEURS error 59.6 → 17.2% since v1.1); `duplexjev` default | DuplexJev-4B-Turn + para LoRA + conversation replay + language LoRAs | 75 | 92.5\* | 86.5 | 81.6 | CC BY-NC 4.0 |
 | [**DuplexJev-4B-Para-LiveTurn**](https://huggingface.co/adventists-ai/DuplexJev-4B-Para-LiveTurn) 🆕 | live end-of-turn for LiveKit / Pipecat agents on ~10 GB (+ accents); eot-bench false cut-offs 25.5% en / 38.8% zh | DuplexJev-4B-Para + accent and pause-level turn data | 76 | 92.0\* | 89.5 | 90.5 | CC BY-NC 4.0 |
-| [**DuplexJev-32B-Stress**](https://huggingface.co/adventists-ai/DuplexJev-32B-Stress) | 32B-Turn + hears stress and pauses (one 80 GB GPU) | DuplexJev-32B-Turn + stress LoRA | **96** | 94.4\* | 92.1 | 91.4 | CC BY-NC 4.0 |
+| [**DuplexJev-32B-Stress**](https://huggingface.co/adventists-ai/DuplexJev-32B-Stress) | 32B-Turn + hears stress and pauses (one 80 GB GPU) | DuplexJev-32B-Turn + stress LoRA | 96 | 94.4\* | 92.1 | 91.4 | CC BY-NC 4.0 |
 | [**DuplexJev-4B-Stress**](https://huggingface.co/adventists-ai/DuplexJev-4B-Stress) | 4B-Turn + stress and pauses (~10 GB; ~4 pts lower on VoiceBench) | DuplexJev-4B-Turn + stress LoRA | 73 | 91.8\* | 89.1 | 91.4 | CC BY-NC 4.0 |
 
 🎙️ **Live end of turn (v1.2):** on LiveKit's [eot-bench](https://github.com/livekit/eot-bench), false cut-offs at a 300 ms budget: 32B-Para v1.2 **20.0%** with the assistant's previous message as the plugins send it, 23.5% audio only (Smart Turn v3.2 35.2%, LiveKit Turn Detector v1 9.9%). Plugins: `pipecat-duplexjev`, `livekit-plugins-duplexjev`.