Update README.md

README.md

@@ -5,7 +5,7 @@ base_model: Qwen/Qwen3-ASR-0.6B-hf
 pipeline_tag: audio-classification
 tags: [turn-detection, end-of-turn, turn-taking, voice-agent, endpointing, semantic-vad, speech]
 model-index:
-- name: voxturn-v1
+- name: voxturn
   results:
   - task: {type: audio-classification, name: End-of-turn detection}
     dataset: {type: livekit/eot-bench-data, name: eot-bench (English, validation)}
@@ -16,19 +16,28 @@ model-index:
     - {type: latency_at_10pct_cutoff, value: 335, name: Latency @ 10% cut-offs (ms)}
 ---
 
-# voxturn v1
+# voxturn
 
-**An open audio turn-taking model that answers several questions about the moment in one forward pass**: has the
+<!-- **An open audio turn-taking model that answers several questions about the moment in one forward pass**: has the
 user finished their turn, how soon will they go on, is this a backchannel or an interruption, is the sound speech at
 all. It hears the last 8 seconds of user audio (optionally with a transcript or the dialog so far), runs on one GPU in
-~8 ms, and is measured on all 14 languages of [eot-bench](https://huggingface.co/datasets/livekit/eot-bench-data).
+~8 ms, and is measured on all 14 languages of [eot-bench](https://huggingface.co/datasets/livekit/eot-bench-data). -->
+
+**An open, audio-conditioned turn-taking model that answers several conversational questions in one forward pass.**
+Has the user finished their turn? When will they resume? Is this feedback or an interruption? Is this a non-speech
+sound? VoxTurn analyzes the last 8 seconds of audio, optionally with a transcript or dialog context. On an NVIDIA
+L40S (BF16, optimized inference), one end-of-turn decision takes **8.3 ms p50**, including feature extraction.
+End-of-turn detection is evaluated in 14 languages on [eot-bench](https://huggingface.co/datasets/livekit/eot-bench-data).
+
+![Figure (a). Conversational timing: mid-turn pauses, completed turns, backchannels, and interruptions call for different agent behavior.](figures/a_conversational_timing.png)
+
 
 | | |
 |---|---|
 | **English** | 11.5% false cut-offs at a 300 ms latency budget — ahead of the commercial Baton (12.3%) and Deepgram Flux (12.9%) |
 | **Multilingual** | 19.9% mean over 14 languages — the best published mean among open models (LiveKit v1-mini 28.5%, Smart Turn v3.2 39.8%) |
 | **One pass, many questions** | 4 questions in 9.0 ms vs 33.2 ms asked one by one (3.7×), answers identical up to rounding |
-| **Size** | one 0.8B model; 8.3 ms per decision, 1.9 GB GPU memory |
+| **Size** | one 0.8B model; 8.3 ms per decision on Nvidia L40S, 1.9 GB GPU memory |
 | **Drop-in for voice agents** | turn analyzers for [Pipecat](#use-with-pipecat) (16.7% of turns cut off vs 45.3% with Smart Turn v3) and [LiveKit Agents](#use-with-livekit-agents) (9.3% vs 18.0% with LiveKit's local turn detector, at lower latency) |
 
 ## Quick start
@@ -131,6 +140,8 @@ recheck_secs=...)`). Create the detector once per worker process (e.g. in `prewa
 
 ## What it answers
 
+![Figure (b). Shared audio prefix and block-isolated question prompting in VoxTurn v1.](figures/b_model_architecture.png)
+
 Every question is a multiple-choice prompt; all of them share one prefix (audio, optional text, who is speaking) and
 are answered in the same forward pass.
 
@@ -152,12 +163,13 @@ and Chinese (`lang`); the audio itself can be in any language.
 | Backbone | [Qwen3-ASR-0.6B](https://huggingface.co/Qwen/Qwen3-ASR-0.6B-hf) (Apache-2.0) |
 | Readout | no text is generated: each question is written out with its options, and a linear head scores the hidden state at each option's last token; softmax over the options |
 | One pass | question blocks follow a shared prefix under a block-isolating attention mask, each at the positions it would have if asked alone — so one pass with k questions equals k separate passes |
-| Files | `voxturn-v1.safetensors` (47 MB), `voxturn.py` (inference), `pipecat_voxturn.py` (Pipecat analyzer), `livekit_voxturn.py` (LiveKit turn detector), `questions.json` (prompts), `config.json` |
 
 ## Evaluation
 
 ### eot-bench (validation, 100 ms minimum silence)
 
+![Figure (d). Task-specific evaluations: eot-bench, Easy-Turn, and SLR99. Lower false cut-off rates are better.](figures/d_benchmarks.png)
+
 Metrics from `eot-harness`: false cut-offs at a latency budget (lower is better) and latency at a cut-off rate. eot-bench
 audio and labels were never used for training.
 
@@ -198,6 +210,8 @@ the 14 languages and has the best mean.
 
 ### Speed (NVIDIA L40S, bf16, 8 s audio)
 
+![Figure (e). Single-pass inference efficiency and separate 150-turn framework replay results.](figures/e_speed_and_deployment.png)
+
 | | p50 | p95 |
 |---|---|---|
 | 1 question (`eot`) | 8.3 ms | 8.4 ms |
@@ -205,10 +219,10 @@ the 14 languages and has the best mean.
 | the same 4 questions as 4 passes | 33.2 ms | 33.3 ms |
 | 4 clips in one batch, 1 question each | 16.0 ms | 16.1 ms |
 
-End to end through `predict`, including feature extraction; GPU memory 1.9 GB. The default `fast=True` merges the
+End-to-end inference time through `predict`, including feature extraction (not end-of-turn response latency); GPU memory 1.9 GB. The default `fast=True` merges the
 adapter weights into the backbone and replays CUDA graphs captured on first use of each input shape (the very first
-call takes ~0.5 s, each new shape ~0.1 s). `VoxTurn.from_pretrained(path, fast=False)` runs the unmerged weights eagerly
-(~38 ms) and reproduces the evaluation computation exactly.
+call takes approximately 0.5 s, each new shape approximately 0.1 s). `VoxTurn.from_pretrained(path, fast=False)` runs the unmerged weights eagerly
+(approximately 38 ms) and reproduces the evaluation computation exactly.
 
 Training details will be described in a technical report.