| --- |
| tags: |
| - speech-dialogue |
| - expressive-speech |
| - text-to-speech |
| - reproducibility |
| --- |
| |
| # EchoLoc |
|
|
| This repository contains the EchoLoc implementation and executable entrypoints for data construction, calibration, Thinker/Talker training and inference, speech rendering, VStyle/render evaluation, benchmark inference, and statistical audits. |
|
|
| Large datasets, generated audio, cached features, job logs, API credentials, and precomputed result tables are intentionally omitted. Provide local paths through command-line arguments, YAML configuration, or environment variables. |
|
|
| ## Released checkpoints |
|
|
| - Thinker PEFT adapter: `checkpoints/thinker/` |
| - Talker checkpoint: `checkpoints/talker/` |
| - Thinker base: `Qwen/Qwen3-Omni-30B-A3B-Instruct` |
| - Talker origins: `Qwen/Qwen3-TTS-12Hz-1.7B-Base` and `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` |
|
|
| Download this private repository, including both checkpoints, with: |
|
|
| ```bash |
| hf download zsy814/EchoLoc --local-dir echoloc-release |
| ``` |
|
|
| The Talker checkpoint is loaded over the Base/VoiceDesign merged origin described in `checkpoints/talker/README.md`. This repository is kept private during double-blind review and can be made public after anonymity restrictions are lifted. |
|
|
| ## Directory map |
|
|
| - `data_construction/`: context/query generation, quality filtering, acoustic controls, oracle targets, and SFT export. |
| - `calibration_agent/`: planner, critics, diagnostic memory, repair loop, and supervisor. |
| - `model/thinker_talker/`: EchoLoc Thinker/Talker task, model, training, and inference code. |
| - `model/qwen_omni_finetune/`: Qwen-Omni fine-tuning and evaluation adapters. |
| - `rendering/universal_tts/`: dual-path rendering framework and synthesizer adapters. |
| - `evaluation/vstyle/`: VStyle prompts, inference, scoring, ASR, and aggregation. |
| - `evaluation/render_metrics/` and `evaluation/esd/`: AutoPCP, emotion embedding, smoothness, WER, ESD alignment, and report generation. |
| - `inference/benchmarks/`: generic benchmark input, Thinker/Talker inference, and output export. |
| - `experiments/`: reviewer-facing statistical analyses. |
| - `checkpoints/thinker/`: best Iter006 Thinker PEFT adapter and tokenizer assets. |
| - `checkpoints/talker/`: best Talker-46000 state dictionary and tokenizer assets. |
| - `CHECKPOINTS.md`: checkpoint provenance, byte sizes, and SHA-256 digests. |
| - `SOURCE_INVENTORY.tsv`: implementation-area inventory. |
| - `MANIFEST.tsv`: size and SHA-256 for every source-release file; checkpoint hashes are recorded separately in `CHECKPOINTS.md`. |
|
|
| Start with `ENVIRONMENT.md`, then follow `RUNBOOK.md`. Run `python tools/validate_archive.py` for a source and manifest integrity check. |
|
|
