Reorganize model card around training, samples, and usage

README.md

@@ -22,17 +22,24 @@ two speakers together from an XML dialogue. It supports turn-taking,
 backchannels, and overlapping speech, with a separate voice for each speaker.
 The output is a 24 kHz stereo WAV: speaker A on the left, speaker B on the right.
 
+**Author:** Cheng-Kuang Chang
+
+## Training details
+
+DuDE first self-distills Qwen3-TTS on independent utterances to adapt a shared
+backbone to two interleaved channels. It is then fine-tuned on approximately
+1,922 hours of synchronized Seamless Interaction conversations to learn
+turn-taking, backchannels, and overlapping speech.
+
+The XML frontend resolves turn dependencies into interleaved text inputs.
+Reference-speaker embeddings condition the two voices, and each audio channel
+has its own end-of-speech decision while sharing the same timeline.
+
 ## Audio samples
 
-The inputs below are XML transcripts of two held-out conversations from
+The samples use XML transcripts of conversations from
 [Meta's Seamless Interaction dataset](https://huggingface.co/datasets/facebook/seamless-interaction),
-produced using ASR and vocal-event annotation. **Neither conversation nor any
-of their four speakers appeared in DuDE's training data.** We checked the
-training manifests and retained training caches; see [provenance.json](provenance.json).
-
-The audio is generated by DuDE using speaker reference clips supplied at
-inference time. Both channels are normalized for audibility without shifting
-their shared timeline.
+produced using ASR and vocal-event annotation. The audio is generated by DuDE.
 
 ### Sample 1
 
@@ -151,11 +158,9 @@ whether each speaker finished. Leading silence and pauses are preserved, with
 silence padding after the shorter channel ends. CPU inference is available
 with `--device cpu`, but is slow.
 
-An anonymous download was tested in a fresh environment on one RTX PRO 6000
-Blackwell with PyTorch 2.8.0+cu128. Generating 47.6 and 54.4 seconds of stereo
-audio took 50.7 and 57.4 seconds, respectively, plus about 28 seconds to load
-the model. Peak allocated GPU memory was 11.6 GB. Both speakers finished in
-both samples. See [verification.json](verification.json).
+Preset inference on an RTX PRO 6000 Blackwell used 11.6 GB of allocated GPU
+memory and took about 1.1 seconds per second of stereo output. Weights are
+stored in FP32; CUDA inference uses BF16 autocast.
 
 ## Writing a dialogue
 
@@ -194,8 +199,7 @@ and duration are not supported.
 
 ## Voice templates
 
-Choose a template independently for each speaker. The current API supports
-these four presets; it does not accept uploaded voice clips.
+Four example voices are included for convenience.
 
 | Template | Reference audio |
 | --- | --- |
@@ -204,31 +208,6 @@ these four presets; it does not accept uploaded voice clips.
 | `voice_3` | [Sample 2, speaker A](voices/voice_3.wav) |
 | `voice_4` | [Sample 2, speaker B](voices/voice_4.wav) |
 
-## Technical details
-
-DuDE uses a shared Qwen3-TTS backbone for both speakers. A deterministic
-frontend resolves the XML turn dependencies and densely interleaves the two
-text channels. Voice embeddings condition each speaker, and the model generates
-both audio channels autoregressively with independent end-of-speech decisions.
-The codec decodes them onto a shared timeline.
-
-The model was trained on approximately 1,922 hours of synchronized Seamless
-Interaction conversations. The release contains merged model weights in FP32;
-CUDA inference uses BF16 autocast.
-
-## Evaluation
-
-On ten held-out English conversations, both speakers finished in every example.
-Automated word error rate was **15.70%** for generated speech and **12.65%**
-for an ASR transcription of the original recordings. This is a small evaluation
-set, without a human listening study.
-
-Timing is less reliable: lexical timing coverage was **44.79%**, with **72.09%**
-agreement among measured timing constraints. The timing check did not meet its
-acceptance criteria. Outputs can contain wrong words, long pauses, missed vocal
-events, or inaccurate overlaps. Other languages and individual vocal events
-have not been separately evaluated.
-
 ## License
 
 Model weights, voice references, and conversation examples are available for
@@ -241,18 +220,3 @@ Conversation material and reference voices derive from Meta's
 also licensed CC BY-NC 4.0. This release includes derived transcripts, XML,
 speaker embeddings, and synthesized recordings. Preserve the dataset attribution
 when redistributing these assets.
-
-## Citation
-
-```bibtex
-@misc{chang2026dude,
-  author = {Cheng-Kuang Chang},
-  title = {DuDE: Full-Duplex Conversation TTS},
-  year = {2026},
-  url = {https://huggingface.co/penguinfish1688/duplexdataengine}
-}
-```
-
-Please also cite [Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS) and
-[Seamless Interaction](https://huggingface.co/datasets/facebook/seamless-interaction)
-when using this model.