Context Spanning · DuetaSpan-7B

[!Note] This repository holds the released DuetaSpan-7B weights described in Context Spanning: A Communication Framework for Full-Duplex Speech Models and External LLM Backends (Go, Kim, Cha, Shin — Mindlogic, 2026, arXiv:2609.33443). The inference and training code lives at mindlogic-ai/Context-Spanning; the weights load only through that package.

A full-duplex speech model that keeps listening and talking while it calls an external backend. When the model emits <ret>, the recent user audio is transcribed, a tool router (with an LLM for knowledge questions) returns one reference, and that reference is written into the model's own context stream as-is as a Context Span block. The frame clock never waits, and the model reads exact values (time, weather, prices, distances) instead of a latent summary.

Files

Mimi and the text tokenizer are taken from the PersonaPlex repository at load time; the moshi package is vendored in the code repository (contextspan/moshi/).

Model Overview

• Type: full-duplex speech-to-speech model (Moshi architecture, PersonaPlex fork)
• Base model: nvidia/personaplex-7b-v1, all trainable parameters fine-tuned
• Parameters: 7B
• Audio codec: Mimi, 12.5 Hz, 8 codebooks per channel (user and agent)
• Frame clock: 80 ms per step
• Sequence convention: <ret> = 4, span open = 12, span close = 13, text pad = 3
• Released checkpoint: 2026-09-23
• Backends used for the results: router google/gemma-4-26B-A4B-it on vLLM, ASR Qwen3-ASR-1.7B
• Language: English

Benchmark Results

Full runs of the released checkpoint with the released router (Gemma-4-26B-A4B on vLLM) and ASR (Qwen3-ASR-1.7B); protocols and scorers are in benchmark/README.md. Rows marked † are reprinted from the original benchmark papers (Full-Duplex-Bench, PersonaPlex) or from MoshiRAG. Arrows give the direction of better.

Table 1. Full-Duplex-Bench v1 (727 clips).

TOR is the take-over rate, the fraction of clips in which the model takes the turn; JSD is the Jensen-Shannon divergence from the human backchannel timing distribution.

Table 2. Full-Duplex-Bench v3 (100 scenarios, 12 mock tools). MoshiRAG is the released kyutai/moshika-rag checkpoint run with its own inference code, its reference LLM replaced by the same tool router and prompt as Ours (router LLM Gemma-3-27B-it).

Tool-use metrics are fractions of the 100 scenarios; rates are percentages; latency is the task-completion latency in seconds.

Table 3. Spoken QA and math reasoning (accuracy, %). ref. = the router's reference answer is judged; resp. = the model's spoken response is judged. LlamaQ 300, WebQ 1000, TriviaQA 1000, HaluEval 1000; math 3,822 (AddSub 395, MultiArith 600, SinglEq 508, SVAMP 1000, GSM8K 1319).

Quickstart

The weights load through the contextspan package; there is no transformers loader.

git clone https://github.com/mindlogic-ai/Context-Spanning && cd Context-Spanning
pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install -e .
export HF_TOKEN=<token with access to nvidia/personaplex-7b-v1 and this repo>

Start the two backends (router LLM and ASR; two GPUs by default, one 96 GB GPU with the flags in the repo README), then:

bash scripts/backends.sh start && source scripts/env.sh

# a wav through the model; out.wav is stereo, L = user, R = agent
python main.py infer --input-wav assets/test/question.wav --output-wav out.wav  \
    --output-text out.json --user-name Priya --user-city Sydney

# live conversation in the browser (microphone needs https or localhost)
python main.py serve --host 0.0.0.0 --port 8080 --token <value>

Any OpenAI-compatible server replaces the router and any POST /transcribe -> {"text"} endpoint replaces the ASR. Servers, GPU layout, vLLM flags and environment variables: docs/BACKENDS.md.

License

• Weights (this repository): NVIDIA Open Model License.
• Voice prompts: main.pt MIT; f*/m* CC0 (Kyutai Unmute Voice Donation).
• Code (mindlogic-ai/Context-Spanning): MIT; contextspan/moshi/ keeps Kyutai's own license files.

Acknowledgements

Fine-tuned from PersonaPlex-7B (NVIDIA), which builds on Moshi and the Mimi codec (Kyutai). The benchmark protocols follow MoshiRAG and Full-Duplex-Bench.