Present DuDE as full-duplex conversation TTS with named samples and verified held-out provenance

.gitattributes

@@ -39,3 +39,5 @@ voices/voice_1.wav filter=lfs diff=lfs merge=lfs -text
 voices/voice_2.wav filter=lfs diff=lfs merge=lfs -text
 voices/voice_3.wav filter=lfs diff=lfs merge=lfs -text
 voices/voice_4.wav filter=lfs diff=lfs merge=lfs -text
+examples/sample_1.wav filter=lfs diff=lfs merge=lfs -text
+examples/sample_2.wav filter=lfs diff=lfs merge=lfs -text

CITATION.cff

@@ -1,7 +1,7 @@
 cff-version: 1.2.0
 message: "If you use DuDE, please cite this model release."
 type: software
-title: "DuDE: XML-conditioned Duplex Text-to-Speech"
+title: "DuDE: Full-Duplex Conversation TTS"
 authors:
   - family-names: Chang
     given-names: Cheng-Kuang

README.md

@@ -4,7 +4,8 @@ language:
 license: cc-by-nc-4.0
 pipeline_tag: text-to-speech
 tags:
-- duplex
+- full-duplex
+- conversation-tts
 - dialogue
 - stereo
 - xml
@@ -12,34 +13,32 @@ tags:
 base_model: Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
 datasets:
 - facebook/seamless-interaction
-model-index:
-- name: DuDE-TTS
-  results: []
 ---
 
-# DuDE: XML-conditioned Duplex Text-to-Speech
+# DuDE: Full-Duplex Conversation TTS
 
-**By Cheng-Kuang Chang.** Generate two synchronized speakers from an anchored
-XML dialogue and two selectable voice templates. The output is a normalized
-24 kHz stereo WAV: **speaker A on the left, speaker B on the right**.
+DuDE is a full-duplex conversation TTS (text-to-speech) model that generates
+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.
 
-Listen to the two featured recordings and read their complete input XML below.
+## Audio samples
 
-This repository contains a standalone Stage III inference release: merged
-model weights, the deterministic XML interleaver, the audio codec, voice
-presets, and inference code. It does not require an external transcript-generating
-language model, training caches, or a separate adapter/base checkpoint.
-
-## Featured examples and complete input XML
+The inputs below are XML transcripts of two held-out 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).
 
-These are the requested checkpoint-8565 recordings, normalized for listening.
-Both channels preserve their shared origin; neither lane is trimmed or shifted.
+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.
 
-### V01_S0192_I00000307-0000-student
+### Sample 1
 
-<audio controls src="https://huggingface.co/penguinfish1688/duplexdataengine/resolve/main/examples/V01_S0192_I00000307-0000-student.wav"></audio>
+<audio controls src="https://huggingface.co/penguinfish1688/duplexdataengine/resolve/main/examples/sample_1.wav"></audio>
 
-[Download stereo audio](examples/V01_S0192_I00000307-0000-student.wav) · [Input XML](examples/V01_S0192_I00000307-0000.xml) · Voices: `voice_1` / `voice_2`
+[Download stereo audio](examples/sample_1.wav) · [Input XML](examples/sample_1.xml) · Voices: `voice_1` / `voice_2`
 
 <details>
 <summary>Show complete input XML</summary>
@@ -73,11 +72,11 @@ Both channels preserve their shared origin; neither lane is trimmed or shifted.
 
 </details>
 
-### V01_S0220_I00000126-0000-student
+### Sample 2
 
-<audio controls src="https://huggingface.co/penguinfish1688/duplexdataengine/resolve/main/examples/V01_S0220_I00000126-0000-student.wav"></audio>
+<audio controls src="https://huggingface.co/penguinfish1688/duplexdataengine/resolve/main/examples/sample_2.wav"></audio>
 
-[Download stereo audio](examples/V01_S0220_I00000126-0000-student.wav) · [Input XML](examples/V01_S0220_I00000126-0000.xml) · Voices: `voice_3` / `voice_4`
+[Download stereo audio](examples/sample_2.wav) · [Input XML](examples/sample_2.xml) · Voices: `voice_3` / `voice_4`
 
 <details>
 <summary>Show complete input XML</summary>
@@ -115,62 +114,56 @@ Both channels preserve their shared origin; neither lane is trimmed or shifted.
 
 ## Download and run
 
-Use Python 3.12 and a CUDA-capable NVIDIA GPU. Install matching PyTorch and
-torchaudio builds for your CUDA environment first. The model uses FP32 master
-weights with BF16 inference autocast, matching the evaluated checkpoint.
+Use Python 3.12 and a CUDA-capable NVIDIA GPU. The download includes the model,
+audio codec, XML frontend, four voice templates, and inference code. Requirements
+pin PyTorch and torchaudio 2.8.0; use the corresponding CUDA builds for your system.
 
 ```bash
 pip install "huggingface_hub>=0.34,<1"
-hf download penguinfish1688/duplexdataengine --local-dir DuDE-TTS
-cd DuDE-TTS
+hf download penguinfish1688/duplexdataengine --local-dir DuDE
+cd DuDE
 pip install -r requirements.txt
 python infer.py \
-  --model . \
-  --xml examples/V01_S0192_I00000307-0000.xml \
+  --model . --xml examples/sample_1.xml \
   --voice-a voice_1 --voice-b voice_2 \
   --seed 20261060 --max-seconds 118 \
   --output dialogue.wav
 ```
 
-Run from Python inside this downloaded directory:
+Or use the Python API from the downloaded directory:
 
 ```python
 from pathlib import Path
 from dude_tts import DuDE
 
-tts = DuDE.from_pretrained(".", device="cuda")
-print(list(tts.voices))
-audio = tts.generate(
-    Path("examples/V01_S0192_I00000307-0000.xml").read_text(),
-    voice_a="voice_1", voice_b="voice_2", seed=20261060, max_seconds=118,
+model = DuDE.from_pretrained(".", device="cuda")
+audio = model.generate(
+    Path("examples/sample_1.xml").read_text(),
+    voice_a="voice_1", voice_b="voice_2",
+    seed=20261060, max_seconds=118,
 )
 audio.save("dialogue.wav")
-print("A/B emitted EOS:", audio.eos)
 ```
 
-The `max_seconds` argument is a safety ceiling, not a desired speaking duration.
-Generation ends when both lanes emit EOS. An unfinished lane at that ceiling
-is reported as `False` in `audio.eos`. Leading silence and pauses are retained;
-the shorter finished lane is zero-padded. CPU execution is available through
-`--device cpu`, but is much slower.
+`max_seconds` is an output limit, not a target speaking duration. Generation
+finishes when both speakers emit an end-of-speech token; `audio.eos` reports
+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.
 
-**Verified from an anonymous public download** in a fresh Python 3.12 environment,
-using PyTorch 2.8.0+cu128 on one RTX PRO 6000 Blackwell. Both featured XML inputs
-generated valid 24 kHz stereo audio and both speakers emitted EOS. The 47.6-second
-and 54.4-second outputs took 50.7 and 57.4 seconds to generate, respectively,
-excluding about 28 seconds to load the model. Peak allocated CUDA memory was
-11.6 GB. The research repository was not installed. See [verification.json](verification.json).
+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).
 
-## How to write the XML
+## Writing a dialogue
 
-Use `<A>` and `<B>` for the two speakers. Exactly **one turn has no `start`**:
-that is the first/root turn. Every other turn needs a `start="e<number>"`
-reference to an inline anchor. Define each anchor exactly once with `<e1/>`,
-`<e2/>`, and so on. The outer `<duplex>` wrapper is optional.
+Use `<A>` and `<B>` for the speakers. Exactly one initial turn has no `start`
+attribute. Every later turn references an inline anchor with `start="e1"`,
+`start="e2"`, and so on. Define each anchor once. The `<duplex>` wrapper is optional.
 
-### Ordinary turn-taking
-
-Put the anchor at the end of the preceding turn:
+For turn-taking, put the anchor at the preceding turn's end:
 
 ```xml
 <duplex>
@@ -180,9 +173,7 @@ Put the anchor at the end of the preceding turn:
 </duplex>
 ```
 
-### Overlap and backchannels
-
-Put the anchor inside a turn to let the other speaker enter there:
+For overlap, put the anchor where the other speaker should enter:
 
 ```xml
 <duplex>
@@ -192,79 +183,76 @@ Put the anchor inside a turn to let the other speaker enter there:
 </duplex>
 ```
 
-Anchors describe dependencies between text positions; they are **not seconds**.
-The static interleaver constructs a dependency graph, removes the XML syntax,
-and places dense text tokens into two alternating lanes. Turn-entry gaps use
-PAD. If two turns on one speaker's lane collide, they are serialized in XML
-order; dependent anchors follow the shifted turn. The TTS model predicts the
-audio timing, so exact overlap and duration are not guaranteed.
+Anchors express turn relationships; the model predicts the audio timing.
+Exact entry times are not guaranteed. Colliding turns for the same speaker
+follow their XML order. All references must exist, and turns must not form a cycle.
 
-Supported vocal events are `<laugh/>`, `<cough/>`, `<breath/>`, `<sigh/>`,
-`<cry/>`, `<sneeze/>`, and `<gasp/>`. Write short spoken acknowledgments such as
-“yeah” and “mm-hmm” as ordinary text. Event quality is not separately measured.
-Use XML escapes such as `&amp;` for literal ampersands. Do not add `type`,
-timestamp, or duration attributes. Undefined or duplicate anchors, cycles,
-multiple roots, unsupported tags, and XML entity declarations are rejected.
+Vocal events: `<laugh/>`, `<cough/>`, `<breath/>`, `<sigh/>`, `<cry/>`,
+`<sneeze/>`, and `<gasp/>`. Write spoken backchannels as ordinary text and
+escape literal ampersands as `&amp;`. Attributes such as `type`, timestamps,
+and duration are not supported.
 
 ## Voice templates
 
-Four frozen reference-voice embeddings are included, two from each featured
-conversation. Choose any preset independently for A and B. Listen to the
-reference clips in `voices/` before choosing. Presets are labeled
-by example and channel, without inferring demographic characteristics.
-The released API accepts these preset IDs; arbitrary uploaded voice cloning is
-not part of this version.
-
-## Model and evaluation
-
-The shared model derives from Qwen3-TTS-12Hz-1.7B-CustomVoice. Stage III trained
-on synchronized Seamless Interaction conversations with our own word/event
-transcripts. The selected corpus contains about 1,922 conversation-hours.
-This release merges rank-128 LoRA into its immutable trained base, including
-the learned channel embeddings. No optimizer state is distributed.
-
-Checkpoint **8565** uses a constant LoRA learning rate of `5e-5`, zero warmup,
-and global batch 8. Training assigns weight 1 to frames overlapping a cached
-word/event plus 320 ms on either side, 0.1 to other frames, and 5 to audio EOS.
-Masks are computed online; no VAD removes audio. Audio EOS follows each lane's
-last annotated word/event plus 240 ms.
-
-On ten held-out conversations, both lanes emitted EOS in **10/10** examples
-(checkpoint 7237: 6/10). Automated generated word error rate was **15.70%**
-(previously 29.63%); the original-audio ASR control was **12.65%**. These are
-small-sample automated measurements, not a general quality guarantee or human
-listening evaluation. Lexical timing coverage was **44.79%**, with **72.09%**
-agreement among measured constraints; the timing acceptance check did not pass.
-Some recordings can still contain wrong words, long pauses, missed events,
-or inaccurate overlaps. English is the evaluated language.
-
-## License and attribution
-
-This model release, reference voice clips/embeddings, and dataset-derived
-examples are provided for **noncommercial use under CC BY-NC 4.0**.
-See [the license](https://creativecommons.org/licenses/by-nc/4.0/).
-The standalone DuDE inference code is Apache-2.0 (`LICENSE-CODE`).
-Qwen code/model/codec components retain their Apache-2.0 notices (`LICENSE-QWEN`).
-
-The reference audio and conversation material derive from
-[Meta's Seamless Interaction dataset](https://huggingface.co/datasets/facebook/seamless-interaction),
-licensed CC BY-NC 4.0. We made transcripts, XML layouts, reference embeddings,
-and generated audio; generated examples are model outputs, not original recordings.
-Preserve this attribution when redistributing these assets. Use voice templates
-responsibly and do not misrepresent synthetic speech as an original recording.
+Choose a template independently for each speaker. The current API supports
+these four presets; it does not accept uploaded voice clips.
+
+| Template | Reference audio |
+| --- | --- |
+| `voice_1` | [Sample 1, speaker A](voices/voice_1.wav) |
+| `voice_2` | [Sample 1, speaker B](voices/voice_2.wav) |
+| `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
+noncommercial use under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
+DuDE inference code is Apache-2.0; see [LICENSE-CODE](LICENSE-CODE).
+Qwen components retain their Apache-2.0 notices in [LICENSE-QWEN](LICENSE-QWEN).
+
+Conversation material and reference voices derive from Meta's
+[Seamless Interaction dataset](https://huggingface.co/datasets/facebook/seamless-interaction),
+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: XML-conditioned Duplex Text-to-Speech},
+  title = {DuDE: Full-Duplex Conversation TTS},
   year = {2026},
-  howpublished = {Hugging Face model release},
   url = {https://huggingface.co/penguinfish1688/duplexdataengine}
 }
 ```
 
-Please also credit [Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS) and
+Please also cite [Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS) and
 [Seamless Interaction](https://huggingface.co/datasets/facebook/seamless-interaction)
-when building on this work.
+when using this model.

dude_tts/interleave.py

@@ -1,4 +1,4 @@
-"""Stage III text frontend: XML -> dependency DAG -> dense, ordered lanes.
+"""Dialogue text frontend: XML -> dependency DAG -> dense, ordered lanes.
 
 No timestamps or XML tokens enter the lanes. Start anchors are token boundaries.
 Same-channel turns are serialized in XML order; a delayed turn carries all of

examples/examples.json

@@ -1,20 +1,24 @@
 [
   {
-    "id": "V01_S0192_I00000307-0000",
-    "xml": "examples/V01_S0192_I00000307-0000.xml",
-    "audio": "examples/V01_S0192_I00000307-0000-student.wav",
+    "id": "sample_1",
+    "xml": "examples/sample_1.xml",
+    "audio": "examples/sample_1.wav",
     "voice_a": "voice_1",
     "voice_b": "voice_2",
     "seed": 20261060,
-    "max_seconds": 118.0
+    "max_seconds": 118.0,
+    "label": "Sample 1",
+    "source_conversation": "V01_S0192_I00000307"
   },
   {
-    "id": "V01_S0220_I00000126-0000",
-    "xml": "examples/V01_S0220_I00000126-0000.xml",
-    "audio": "examples/V01_S0220_I00000126-0000-student.wav",
+    "id": "sample_2",
+    "xml": "examples/sample_2.xml",
+    "audio": "examples/sample_2.wav",
     "voice_a": "voice_3",
     "voice_b": "voice_4",
     "seed": 20261063,
-    "max_seconds": 134.0
+    "max_seconds": 134.0,
+    "label": "Sample 2",
+    "source_conversation": "V01_S0220_I00000126"
   }
 ]

provenance.json

@@ -0,0 +1,85 @@
+{
+  "passed": true,
+  "checked_at_utc": "2026-09-27T02:40:28.055408+00:00",
+  "scope": "DuDE training data; upstream Qwen pretraining coverage is not known",
+  "examples": [
+    {
+      "label": "Sample 1",
+      "conversation_id": "V01_S0192_I00000307",
+      "example_id": "V01_S0192_I00000307-0000",
+      "split": "dev",
+      "participants": {
+        "A": "V01_P1356",
+        "B": "V01_P1357"
+      },
+      "transcript_source": "own_Qwen3-ASR+ForcedAligner+PANNs",
+      "dataset_transcript_downloaded": false,
+      "annotation_sha256": "76474198f55a6d7afd7f826cd910b9841af8293a32413f2f8b672c1ef7fd4996",
+      "public_example_id": "sample_1"
+    },
+    {
+      "label": "Sample 2",
+      "conversation_id": "V01_S0220_I00000126",
+      "example_id": "V01_S0220_I00000126-0000",
+      "split": "dev",
+      "participants": {
+        "A": "V01_P1409",
+        "B": "V01_P1500"
+      },
+      "transcript_source": "own_Qwen3-ASR+ForcedAligner+PANNs",
+      "dataset_transcript_downloaded": false,
+      "annotation_sha256": "d3d434ee2d09aa3916f1883df7e1b1134125b285a2a09cac64171c49110d1cb6",
+      "public_example_id": "sample_2"
+    }
+  ],
+  "training_manifests": [
+    {
+      "corpus": "seamless-100h",
+      "train_conversations": 1468,
+      "sha256": "05c7e77c69078b9d68ef3c30a68392d893f82f7a5cca94bdbf2e3a4ced40d282",
+      "conversation_matches": [],
+      "participant_matches": [],
+      "featured_conversations_in_dev": [
+        "V01_S0220_I00000126"
+      ]
+    },
+    {
+      "corpus": "seamless-500h-v1",
+      "train_conversations": 7365,
+      "sha256": "2967b0c49574182fd14d56571107dbc264f199d6cd3ed39ea71bc9956a81203a",
+      "conversation_matches": [],
+      "participant_matches": [],
+      "featured_conversations_in_dev": [
+        "V01_S0192_I00000307",
+        "V01_S0220_I00000126"
+      ]
+    },
+    {
+      "corpus": "seamless-2000h-v1",
+      "train_conversations": 31493,
+      "sha256": "08d437e07c00ebd4a73feed219e0048e69d1bdcb681dac12f20b0fd9c803758e",
+      "conversation_matches": [],
+      "participant_matches": [],
+      "featured_conversations_in_dev": [
+        "V01_S0192_I00000307",
+        "V01_S0220_I00000126"
+      ]
+    }
+  ],
+  "training_codec_caches": [
+    {
+      "corpus": "conversation-2000h-v1",
+      "shards": 8,
+      "examples": 30854,
+      "featured_conversation_matches": []
+    },
+    {
+      "corpus": "conversation-500h-full-v1",
+      "shards": 4,
+      "examples": 7190,
+      "featured_conversation_matches": []
+    }
+  ],
+  "transcript_note": "XML uses our ASR/event annotations of held-out Seamless Interaction audio, not dataset-provided transcripts.",
+  "voice_reference_note": "Speaker reference clips are provided during inference; they are not training examples."
+}

release.json

@@ -1,7 +1,6 @@
 {
-  "format": "dude_stage3_merged_v1",
+  "format": "dude_duplex_v1",
   "version": "1.0.0",
-  "checkpoint_step": 8565,
   "author": "Cheng-Kuang Chang",
   "base_model": "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
   "base_revision": "0c0e3051f131929182e2c023b9537f8b1c68adfe",
@@ -25,22 +24,26 @@
   },
   "examples": [
     {
-      "id": "V01_S0192_I00000307-0000",
-      "xml": "examples/V01_S0192_I00000307-0000.xml",
-      "audio": "examples/V01_S0192_I00000307-0000-student.wav",
+      "id": "sample_1",
+      "xml": "examples/sample_1.xml",
+      "audio": "examples/sample_1.wav",
       "voice_a": "voice_1",
       "voice_b": "voice_2",
       "seed": 20261060,
-      "max_seconds": 118.0
+      "max_seconds": 118.0,
+      "label": "Sample 1",
+      "source_conversation": "V01_S0192_I00000307"
     },
     {
-      "id": "V01_S0220_I00000126-0000",
-      "xml": "examples/V01_S0220_I00000126-0000.xml",
-      "audio": "examples/V01_S0220_I00000126-0000-student.wav",
+      "id": "sample_2",
+      "xml": "examples/sample_2.xml",
+      "audio": "examples/sample_2.wav",
       "voice_a": "voice_3",
       "voice_b": "voice_4",
       "seed": 20261063,
-      "max_seconds": 134.0
+      "max_seconds": 134.0,
+      "label": "Sample 2",
+      "source_conversation": "V01_S0220_I00000126"
     }
   ],
   "license": "cc-by-nc-4.0"

verification.json

@@ -5,12 +5,11 @@
   "anonymous_download": true,
   "isolated_environment": true,
   "research_checkout_required": false,
-  "source_checkpoint_step": 8565,
   "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
   "torch": "2.8.0+cu128",
   "examples": [
     {
-      "id": "V01_S0192_I00000307-0000",
+      "id": "sample_1",
       "eos": [
         true,
         true
@@ -25,10 +24,11 @@
       "correlation_with_featured_audio": [
         0.9999125589163387,
         0.9996439714705407
-      ]
+      ],
+      "label": "Sample 1"
     },
     {
-      "id": "V01_S0220_I00000126-0000",
+      "id": "sample_2",
       "eos": [
         true,
         true
@@ -43,7 +43,8 @@
       "correlation_with_featured_audio": [
         0.9999395525329423,
         0.9998978182617578
-      ]
+      ],
+      "label": "Sample 2"
     }
   ],
   "peak_gpu_memory_gb": 11.553627648,
@@ -58,5 +59,6 @@
     "Both lanes emit EOS",
     "Finite normalized 24 kHz stereo waveform",
     "Shared clock preserved; waveform correlation above 0.99 with each featured lane"
-  ]
+  ],
+  "model_version": "1.0.0"
 }

voices/presets.json

@@ -1,6 +1,6 @@
 {
   "voice_1": {
-    "label": "Example 1 \u00b7 speaker A",
+    "label": "Sample 1 \u00b7 speaker A",
     "embedding": [
       0.0054931640625,
       0.0274658203125,
@@ -2060,7 +2060,7 @@
     "license": "cc-by-nc-4.0"
   },
   "voice_2": {
-    "label": "Example 1 \u00b7 speaker B",
+    "label": "Sample 1 \u00b7 speaker B",
     "embedding": [
       0.103515625,
       -0.007080078125,
@@ -4120,7 +4120,7 @@
     "license": "cc-by-nc-4.0"
   },
   "voice_3": {
-    "label": "Example 2 \u00b7 speaker A",
+    "label": "Sample 2 \u00b7 speaker A",
     "embedding": [
       -0.0400390625,
       0.0390625,
@@ -6180,7 +6180,7 @@
     "license": "cc-by-nc-4.0"
   },
   "voice_4": {
-    "label": "Example 2 \u00b7 speaker B",
+    "label": "Sample 2 \u00b7 speaker B",
     "embedding": [
       0.0634765625,
       -0.058837890625,