feat: Qwen3-TTS 0.6B voice clone

.github/workflows/README.md

@@ -0,0 +1,8 @@
+# GitHub Actions Workflows
+
+| Workflow | Trigger | Description |
+|----------|---------|-------------|
+| [CI](ci.yml) | Push to `main`, PRs, manual | Runs `cargo fmt`, `clippy`, tests, and doc build |
+| [Release-plz](release-plz.yml) | Push to `main`, manual | Automates crate releases to crates.io and opens release PRs |
+| [Update benchmark table](update-bench.yml) | Push to `main` (when `bench/results/**.csv` changes), manual | Regenerates the benchmark table in README.md |
+| [Export ONNX](export-onnx.yml) | Manual | Exports Qwen3-TTS to ONNX (FP32 + INT4), validates, and uploads to HuggingFace. Supports `voicedesign` (1.7B) and `clone` (0.6B Base) variants via input selector. |

.github/workflows/export-onnx.yml

@@ -3,13 +3,20 @@ name: Export ONNX & Publish to HuggingFace
 on:
   workflow_dispatch:
     inputs:
+      variant:
+        description: "Model variant to export"
+        default: "voicedesign"
+        type: choice
+        options:
+          - voicedesign
+          - clone
       model_id:
-        description: "HuggingFace source model"
-        default: "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign"
+        description: "HuggingFace source model (leave default for selected variant)"
+        default: ""
         type: string
       hf_repo:
-        description: "Target HuggingFace repo (org/name)"
-        default: "wavekat/Qwen3-TTS-1.7B-VoiceDesign-ONNX"
+        description: "Target HuggingFace repo (leave default for selected variant)"
+        default: ""
         type: string
       runner:
         description: "GitHub Actions runner label"
@@ -20,7 +27,7 @@ on:
         default: false
         type: boolean
       revision:
-        description: "HuggingFace revision/branch to push to (e.g. v1, 2026-04-06). Defaults to main."
+        description: "HuggingFace revision/branch to push to. Defaults to main."
         default: ""
         type: string
 
@@ -29,6 +36,11 @@ jobs:
     runs-on: ${{ inputs.runner }}
     timeout-minutes: 120
 
+    env:
+      MODEL_ID: ${{ inputs.model_id != '' && inputs.model_id || (inputs.variant == 'clone' && 'Qwen/Qwen3-TTS-12Hz-0.6B-Base' || 'Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign') }}
+      HF_REPO: ${{ inputs.hf_repo != '' && inputs.hf_repo || (inputs.variant == 'clone' && 'wavekat/Qwen3-TTS-0.6B-Base-ONNX' || 'wavekat/Qwen3-TTS-1.7B-VoiceDesign-ONNX') }}
+      OUTPUT_DIR: ${{ inputs.variant == 'clone' && './output/qwen3-tts-0.6b-base' || './output/qwen3-tts-1.7b-voicedesign' }}
+
     steps:
       - name: Free disk space
         run: |
@@ -61,12 +73,23 @@ jobs:
         working-directory: tools/qwen3-tts-onnx
         run: make venv
 
-      - name: Export FP32 ONNX models
+      - name: Export FP32 ONNX models (VoiceDesign)
+        if: inputs.variant == 'voicedesign'
         working-directory: tools/qwen3-tts-onnx
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
-          make export MODEL_ID="${{ inputs.model_id }}"
+          make export MODEL_ID="${{ env.MODEL_ID }}"
+          echo "=== Disk after export ==="
+          df -h /
+
+      - name: Export FP32 ONNX models (Clone — 6 components)
+        if: inputs.variant == 'clone'
+        working-directory: tools/qwen3-tts-onnx
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          make clone-export CLONE_MODEL_ID="${{ env.MODEL_ID }}"
           echo "=== Disk after export ==="
           df -h /
 
@@ -75,25 +98,40 @@ jobs:
         working-directory: tools/qwen3-tts-onnx
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
-        run: make validate MODEL_ID="${{ inputs.model_id }}"
+        run: make validate MODEL_ID="${{ env.MODEL_ID }}" OUTPUT_DIR="${{ env.OUTPUT_DIR }}"
 
       - name: Free PyTorch model cache
         run: |
           rm -rf ~/.cache/huggingface/hub
           echo "=== Disk after cache cleanup ==="
           df -h /
 
-      - name: Quantize to INT4
+      - name: Quantize to INT4 (VoiceDesign)
+        if: inputs.variant == 'voicedesign'
         working-directory: tools/qwen3-tts-onnx
         run: |
           make quantize
           echo "=== Disk after quantize ==="
           df -h /
 
-      - name: Package for HuggingFace
+      - name: Quantize to INT4 (Clone — encoders stay FP32)
+        if: inputs.variant == 'clone'
+        working-directory: tools/qwen3-tts-onnx
+        run: |
+          make clone-base-preset CLONE_MODEL_ID="${{ env.MODEL_ID }}"
+          echo "=== Disk after quantize ==="
+          df -h /
+
+      - name: Package for HuggingFace (VoiceDesign)
+        if: inputs.variant == 'voicedesign'
         working-directory: tools/qwen3-tts-onnx
         run: make hf
 
+      - name: Package for HuggingFace (Clone)
+        if: inputs.variant == 'clone'
+        working-directory: tools/qwen3-tts-onnx
+        run: make clone-hf
+
       - name: Upload to HuggingFace
         working-directory: tools/qwen3-tts-onnx
         env:
@@ -102,25 +140,26 @@ jobs:
           .venv/bin/pip install -q huggingface_hub
           REVISION="${{ inputs.revision }}"
           .venv/bin/huggingface-cli upload \
-            "${{ inputs.hf_repo }}" \
-            ./output/qwen3-tts-1.7b-voicedesign/ \
+            "${{ env.HF_REPO }}" \
+            ${{ env.OUTPUT_DIR }}/ \
             --repo-type model \
-            --commit-message "export: ${{ inputs.model_id }} (run #${{ github.run_number }})" \
+            --commit-message "export: ${{ env.MODEL_ID }} (run #${{ github.run_number }})" \
             ${REVISION:+--revision "$REVISION"}
 
       - name: Job summary
         if: always()
         run: |
           REVISION="${{ inputs.revision }}"
           REVISION_DISPLAY="${REVISION:-main}"
-          HF_URL="https://huggingface.co/${{ inputs.hf_repo }}/tree/${REVISION_DISPLAY}"
+          HF_URL="https://huggingface.co/${{ env.HF_REPO }}/tree/${REVISION_DISPLAY}"
           cat >> "$GITHUB_STEP_SUMMARY" <<EOF
           ## ONNX Export Summary
 
           | | |
           |---|---|
-          | **Source model** | \`${{ inputs.model_id }}\` |
-          | **HuggingFace repo** | [${{ inputs.hf_repo }}](${HF_URL}) |
+          | **Variant** | \`${{ inputs.variant }}\` |
+          | **Source model** | \`${{ env.MODEL_ID }}\` |
+          | **HuggingFace repo** | [${{ env.HF_REPO }}](${HF_URL}) |
           | **Revision** | \`${REVISION_DISPLAY}\` |
           | **Run** | [#${{ github.run_number }}](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}) |
           | **Validation** | ${{ inputs.skip_validation && 'skipped' || 'passed' }} |

.gitignore

@@ -6,3 +6,5 @@ Cargo.lock
 dataset/
 models/
 *.wav
+!tools/qwen3-tts-onnx/fixtures/*.wav
+.claude/

README.md

@@ -20,7 +20,8 @@ Same pattern as
 
 | Backend | Feature flag | Status | License |
 |---------|-------------|--------|---------|
-| [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS) | `qwen3-tts` | ✅ Available | Apache 2.0 |
+| [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS) (VoiceDesign 1.7B) | `qwen3-tts` | ✅ Available | Apache 2.0 |
+| [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS) (Voice Clone 0.6B) | `qwen3-tts` | ✅ Available | Apache 2.0 |
 | [CosyVoice](https://github.com/FunAudioLLM/CosyVoice) | `cosyvoice` | 🚧 Planned | Apache 2.0 |
 
 ## Model weights
@@ -29,39 +30,67 @@ ONNX-converted weights are published under the [`wavekat`](https://huggingface.c
 
 | Backend | Repository | Precision |
 |---------|------------|-----------|
-| Qwen3-TTS | [wavekat/Qwen3-TTS-1.7B-VoiceDesign-ONNX](https://huggingface.co/wavekat/Qwen3-TTS-1.7B-VoiceDesign-ONNX) | FP32, INT4 |
+| Qwen3-TTS VoiceDesign | [wavekat/Qwen3-TTS-1.7B-VoiceDesign-ONNX](https://huggingface.co/wavekat/Qwen3-TTS-1.7B-VoiceDesign-ONNX) | FP32, INT4 |
+| Qwen3-TTS Voice Clone | [wavekat/Qwen3-TTS-0.6B-Base-ONNX](https://huggingface.co/wavekat/Qwen3-TTS-0.6B-Base-ONNX) | FP32, INT4 |
 
 ## Quick start
 
 ```sh
 cargo add wavekat-tts --features qwen3-tts
 ```
 
+### VoiceDesign (prompt-based styling)
+
 ```rust
 use wavekat_tts::{TtsBackend, SynthesizeRequest};
-use wavekat_tts::backends::qwen3_tts::{Qwen3Tts, ModelConfig, ModelPrecision, ExecutionProvider};
-
-// Auto-downloads INT4 model files on first run, runs on CPU (default):
-let tts = Qwen3Tts::new()?;
-
-// Or FP32 on CPU:
-// let tts = Qwen3Tts::from_config(ModelConfig::default().with_precision(ModelPrecision::Fp32))?;
-
-// Or INT4 from a local directory on CUDA:
-// let tts = Qwen3Tts::from_config(
-//     ModelConfig::default()
-//         .with_dir("models/qwen3-tts-1.7b")
-//         .with_execution_provider(ExecutionProvider::Cuda),
-// )?;
+use wavekat_tts::backends::qwen3_tts::Qwen3Tts;
+// use wavekat_tts::backends::qwen3_tts::{ModelConfig, ModelPrecision, ExecutionProvider};
+
+fn main() {
+    let tts = Qwen3Tts::new().unwrap(); // auto-downloads INT4 model on first run
+
+    // For FP32 or GPU inference, use from_config:
+    // let config = ModelConfig::default()
+    //     .with_precision(ModelPrecision::Fp32)       // FP32 instead of INT4
+    //     .with_execution_provider(ExecutionProvider::Cuda); // GPU via CUDA
+    // let tts = Qwen3Tts::from_config(config).unwrap();
+
+    let request = SynthesizeRequest::new("Hello, world")
+        .with_instruction("Speak naturally and clearly.");
+    let audio = tts.synthesize(&request).unwrap();
+    audio.write_wav("output.wav").unwrap();
+    println!("Wrote output.wav ({:.2}s)", audio.duration_secs());
+}
+```
 
-let request = SynthesizeRequest::new("Hello, world")
-    .with_instruction("Speak naturally and clearly.");
-let audio = tts.synthesize(&request)?;
+### Voice Clone (reference-audio cloning)
 
-// Save to WAV (wavekat-core includes WAV I/O via the `wav` feature):
-audio.write_wav("output.wav")?;
+> **Requires a reference WAV file** (`ref.wav`) — a short mono clip of the voice
+> you want to clone, plus a transcript of what is spoken in the clip.
 
-println!("{}s at {} Hz", audio.duration_secs(), audio.sample_rate());
+```rust
+use wavekat_tts::AudioFrame;
+use wavekat_tts::backends::qwen3_tts::{Qwen3TtsClone, CloneRequest};
+// use wavekat_tts::backends::qwen3_tts::{ModelConfig, ModelPrecision};
+
+fn main() {
+    let ref_audio = AudioFrame::from_wav("ref.wav").unwrap(); // 24 kHz mono WAV
+    let tts = Qwen3TtsClone::new().unwrap(); // auto-downloads 0.6B Base INT4 model
+
+    // For FP32 precision:
+    // let config = ModelConfig::default().with_precision(ModelPrecision::Fp32);
+    // let tts = Qwen3TtsClone::from_config(config).unwrap();
+
+    let req = CloneRequest::new(
+        "Text to say in the cloned voice",
+        ref_audio.samples(),
+        24000,
+        "Transcript of the reference clip.",
+    ).with_language("en");
+    let audio = tts.synthesize_clone(&req).unwrap();
+    audio.write_wav("clone_output.wav").unwrap();
+    println!("Wrote clone_output.wav ({:.2}s)", audio.duration_secs());
+}
 ```
 
 Model files are cached by the HF Hub client at `$HF_HOME/hub/` (default `~/.cache/huggingface/hub/`).
@@ -92,10 +121,17 @@ Two trait families:
 Generate a WAV file from text (model files are auto-downloaded on first run):
 
 ```sh
+# VoiceDesign (1.7B)
 cargo run --example synthesize --features qwen3-tts -- "Hello, world\!"
 cargo run --example synthesize --features qwen3-tts -- --instruction "Speak in a warm, friendly tone." "Give every small business the voice of a big one."
-cargo run --example synthesize --features qwen3-tts -- --precision fp32 "Hello"
-cargo run --example synthesize --features qwen3-tts -- --model-dir /path/to/model --output hello.wav "Hello"
+# cargo run --example synthesize --features qwen3-tts -- --precision fp32 "Hello, world\!"
+
+# Voice Clone (0.6B)
+cargo run --example synthesize_clone --features qwen3-tts -- \
+  --ref-audio ref.wav --ref-text "Transcript of the reference clip." \
+  "Text to synthesize in the cloned voice."
+# cargo run --example synthesize_clone --features qwen3-tts -- --precision fp32 \
+#   --ref-audio ref.wav --ref-text "Transcript." "Text to synthesize."
 ```
 
 ## Performance

crates/wavekat-tts/Cargo.toml

@@ -13,7 +13,7 @@ categories = ["multimedia::audio"]
 default = []
 
 # Local inference backends (all ONNX-based)
-qwen3-tts = ["dep:ort", "dep:ndarray", "dep:tokenizers", "dep:npyz", "dep:rand", "dep:hf-hub"]
+qwen3-tts = ["dep:ort", "dep:ndarray", "dep:tokenizers", "dep:npyz", "dep:rand", "dep:hf-hub", "dep:realfft"]
 cosyvoice = ["dep:ort", "dep:ndarray"]
 
 # Execution providers — composable with any ONNX backend feature
@@ -22,7 +22,7 @@ cuda     = ["ort?/cuda"]
 tensorrt = ["ort?/tensorrt"]
 
 [dependencies]
-wavekat-core = { version = "0.0.5", features = ["wav"] }
+wavekat-core = { version = "0.0.7", features = ["wav"] }
 thiserror = "2"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
@@ -34,11 +34,16 @@ tokenizers = { version = "0.21", optional = true, default-features = false, feat
 npyz = { version = "0.8", optional = true }
 rand = { version = "0.9", optional = true }
 hf-hub = { version = "0.5", optional = true, default-features = false, features = ["ureq"] }
+realfft = { version = "3", optional = true }
 
 [[example]]
 name = "synthesize"
 required-features = ["qwen3-tts"]
 
+[[example]]
+name = "synthesize_clone"
+required-features = ["qwen3-tts"]
+
 [[example]]
 name = "bench_rtf"
 required-features = ["qwen3-tts"]

crates/wavekat-tts/examples/synthesize.rs

@@ -259,4 +259,5 @@ fn synthesize_one(
     audio.write_wav(output).expect("failed to write WAV");
 
     eprintln!("Wrote {}", output.display());
+    eprintln!("Done.");
 }