Local Claude Code state (settings.local.json, worktrees/) shouldn't be tracked.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

API shape, encoder precision, and HF repo policy moved from open questions
to confirmed decisions.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Adds two new export scripts plus Makefile targets for the 0.6B Base voice
clone variant:

- export_speaker_encoder.py: ECAPA-TDNN, mel (B,T,128) -> embed (B,1024)
- export_tokenizer_encoder.py: Mimi encoder, waveform (B,T) -> codes (B,16,F)

Talker / code predictor / vocoder / embeddings reuse the existing config-driven
scripts; the new clone-* Makefile targets pin them to the 0.6B model id and a
separate output directory so 1.7B and 0.6B builds don't overlap.

Encoders ship FP32 only (locked in plan); quantize_int4 is unchanged.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Remove help-clone; all clone-* targets already appear in make help via the
grep pattern. Add CLONE_MODEL_ID/CLONE_OUTPUT_DIR to the variables section.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Mimi encoder's conv padding calls .item() which creates data-dependent
guards incompatible with torch.export/dynamo. Switch to the legacy JIT
tracer with a fixed canonical size (10s @ 24kHz = 240k samples).

Host code zero-pads reference audio to 240k samples before inference and
trims output codes based on the original sample count.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Mimi's encoder_transformer uses create_causal_mask with torch.vmap,
which is incompatible with JIT tracing (same issue as the talker).
Apply the existing mask_patch before tracing.

Also add [N/6] step labels to clone-export for clearer progress.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The Mimi model imports create_causal_mask directly into its local
namespace (`from transformers.masking_utils import create_causal_mask`),
so the existing module-level patch doesn't reach it. Add an explicit
patch of `transformers.models.mimi.modeling_mimi.create_causal_mask`.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

generate_clone_onnx.py chains all 6 ONNX models to produce cloned speech
from a reference audio + transcript. No PyTorch at inference time.

Pipeline: ref audio → tokenizer encoder (codes) + speaker encoder (mel →
embed) → ICL prefill → talker decode loop → code predictor → vocoder
→ trim ref portion → output WAV.

Also adds `make clone-generate` target.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Generate a 7.4s reference clip via macOS TTS (say) at 24kHz mono.
Stored at tools/qwen3-tts-onnx/fixtures/ref_clone.wav with gitignore
exceptions in both root and tools-level .gitignore files.

Makefile clone-generate now defaults to this fixture.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

make clone-fixture uses macOS say + librosa to regenerate
fixtures/ref_clone.wav from CLONE_REF_TEXT.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace macOS say-generated fixture with one produced by our own 1.7B
VoiceDesign backend (Apache-2.0 licensed Qwen3-TTS output). The
clone-fixture Makefile target now invokes cargo run --example synthesize.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Use "Speak in a warm and friendly female voice with natural intonation"
instead of the flat "Speak naturally and clearly" for a more natural
sounding reference clip.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Reference fixture now says "Give every small business the voice of a
big one." — the WaveKat tagline, consistent with all other examples.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Regenerate ref_clone.wav with two sentences (6.95s, well above the 3s
minimum for voice cloning) using FP32 precision for best quality.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Match how the committed fixture was actually generated.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Adds sibling Qwen3TtsClone struct with full ONNX pipeline:
mel spectrogram, speaker encoder, tokenizer encoder, ICL prefill,
talker decode loop, code predictor, and vocoder with ref-code trim.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The speaker_encoder.onnx and tokenizer_encoder.onnx use external
.data files for weights, but these were missing from the download
manifest. Also apply prepare_onnx_dir to the model root so ORT
can resolve HF Hub symlinks for the encoder sessions.

- Add .onnx.data entries to CLONE_SHARED_FILES in download.rs
- Skip directories in prepare_onnx_dir (safe for model root)
- Load encoders from symlink-resolved root directory
- Update voice clone doc to reflect implemented status
- Add fixtures/README.md documenting ref_clone.wav

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Load ref.wav before model in clone example to fail early, add ref.wav
requirement note, and print output info after synthesis in all examples.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>