{ "baseline": { "repo": "mlx-community/chatterbox-multilingual-v3", "revision": "03565773edd72e949572557597af8063bb49a18a", "checkpoint_sha256": "e702f2c441e040bd360e59a86c85d462539759d665e41cfd1938adadd74187a3" }, "q8": { "repo": "groxaxo/chatterbox-multilingual-MLX-Q8-G64", "revision": "ca7cd6a1d72b82aff6e76a037a4fb89626b7f9e6", "checkpoint_sha256": "aa938735e36edfa807531d274baa608ec168f8f06ad6c835a3c4071effcf172e", "loader_sha256": "8c976e357b573d8940d0ff54fb0fa7e4b16804e762ba98de69e15013a1acbcff" }, "q6": { "repo": "groxaxo/chatterbox-multilingual-MLX-Q6-G64", "revision": "dc7ce4b14e534fadff82f0908d8114729e196de7", "checkpoint_sha256": "392f8758a072a70a44ec5c70351a9502877880a196377c74d90cf006a8f06c8c", "loader_sha256": "ca33fe5ae6effa0129898cc6e92310d2070ba4089872e82273c033511953cc31" }, "external_tokenizer": { "repo": "mlx-community/S3TokenizerV2", "revision": "e0c9886f0e1c35ae85b1f27277416fb19fc72bec", "checkpoint_sha256": "928726bc1f206a613d36b8f49e297eae9c5593a21bf9b92ddfe2c23f85eb92cc" }, "quantizer": { "repo": "jundot/omlx", "revision": "70f44cfd652d2013dc0f9776dca7518e32101b00", "recipe": "Native affine Q8/Q6, G64, no calibration. Main 651 matrices plus 37 external runtime matrices; unsupported tensors remain FP32." }, "decoder": { "n_timesteps": 10, "implementation": "S3Token2Wav.__call__(speech_tokens, ref_dict, finalize=True); CausalMaskedDiffWithXvec defaults to 10 steps, no override in constructor or call." }, "installed_library_files_sha256": { "chatterbox.py": "3b5f2cf6c422aa8e1e291e62fea86ab10ba4cb0ae78771f7741de056e7c775ee", "s3gen/s3gen.py": "447d22074d026c9f6f1ea19f5c1fccbf037802b715c4fbb70ce251de853afbd4", "s3gen/flow.py": "f790e91c9cc60c3413e52a0d6a112b41d6be02087f1a6475dbdb21e833b7d376" }, "source_protocol_files_sha256": { "run.sh": "80e62c00e7fb44d8af1814be5267cf104ee02a0ff2a56bfbc493cb09686f0d31", "README.md": "f67f4ec31cea9f81859247c4ede698963f670151e384ceaccb7ceadebad7a893", "environment.json": "261cbd6e1aac98ce22031fca805fbaaf0458db1488d5f7eeae7d76541b739152", "requirements-frozen.txt": "890dd68be72aac1ac861862269d34d50315571c9fc775884d052d3293d449789", "docs/original-methods.md": "b8a34c323fe862f8daea68fae3426a58cfda32b98d19d8206873cc6293d485a6", "scripts/benchmark.py": "a614dafa49421c33c1e2b2e7416dbb3c3a880658a065ae81092ee062de24d0bf", "scripts/audit_weights.py": "487d5ed3d569246edb690afa0734cfa1e7ea3049cb2b864968f9863785d20323", "scripts/measure_similarity.py": "c4cf146182523a2e2b7a2859be31ef6b96c81f36dc11fde15599661edb91fc05", "scripts/decoder_fidelity.py": "383c9fe4863fe91558d06a304e8e03876cdd0e5911a731c2f0194c908743ea06", "data/asr-backend-comparison.json": "b269ef366f47da62db09b5526bb59e660c72087381b60c64175d2a0b58f83194", "data/reproducibility.json": "d86fcf6d62b564e1700269a71d1559e2a3dca47d26dd110831ac45852f1b2d47" }, "privacy": "Reference waveform, generated waveforms, voice embeddings, speech-token arrays and decoder-conditioning safetensors remain private. Only derived metrics and hashes are archived.", "version_scope": "Package versions and library file hashes describe this measured environment; source Turbo used a different runtime. No cross-family ranking." }