Parakeet TDT 0.6B v3, ternary (Core ML and ONNX)

Speech recognition: Parakeet TDT 0.6B v3 with ternary encoder weights, as a Core ML model set and an ONNX model set.

• Credit: NVIDIA, parakeet-tdt-0.6b-v3. Moondream, Parakeet Redux ternary weights. Redistributed under CC BY 4.0.

Layout

• coreml/: Core ML compiled models (Preprocessor, Encoder with 2-bit weights, Decoder, JointDecision), with config.json, tokens.txt and SHA256SUMS. macOS 15 or later.
• onnx/: ONNX encoder (2-bit MatMulNBits), decoder, joiner and voice activity model, with config.json, tokens.txt and SHA256SUMS.

Measured accuracy and speed

Core ML model set on the GPU, on a 62.9-minute subset of the AMI meeting corpus: 22.49% word error rate, about 126x real time on an Apple M3 Max.

Files

File	Bytes	SHA-256
config.json	119	8a8d7725b0bb3127eef2168a42c797b97a29adee4f1c0879dd04597dc1871094
coreml/Decoder.mlmodelc/analytics/coremldata.bin	243	f5c3836cf3c871317e6240a47726f3a866bbafa07c58d9d89fb8b675c71a116b
coreml/Decoder.mlmodelc/coremldata.bin	636	61ab80717a58bc216aec2f5baf1f2ff6420341c725597d2972174fc9b77183da
coreml/Decoder.mlmodelc/metadata.json	3223	1881f965785e90ca4c0ed964559bd9f3455b7d5f6140e970e6f4f01c0e032124
coreml/Decoder.mlmodelc/model.mil	11543	85543f13c2a8cc545854d4ddeb377fb3dc871655e2d32e5820b5adfbfcedd559
coreml/Decoder.mlmodelc/weights/weight.bin	23604992	6c6c88c23ee5492a6b229e8cc636ddadb2b01ae0dd559d9cb63123bebf507d77
coreml/Encoder.mlmodelc/analytics/coremldata.bin	243	e1d91863f0931d56db3fba0b6e7e5d638f78538d61182a06a01c512b132c19c6
coreml/Encoder.mlmodelc/coremldata.bin	542	c158fd6bd0d47907f839ef8bc0c1c0fb39f520eedaa0a6a8eb698f867cbff59a
coreml/Encoder.mlmodelc/metadata.json	2746	32802d3d067d216f9926c5f69b1f232c8223912ac2df735f90346b84dd89132b
coreml/Encoder.mlmodelc/model.mil	598228	4382aeb884dfe5a26566d8080136346ac7e27d9ee754989d5d845a5ac45145fe
coreml/Encoder.mlmodelc/weights/weight.bin	209137408	b1bc0a0d9d3cb333bfa94d2c61723a9e3729232f6548c25b6c9c8cc80cc2f845
coreml/JointDecision.mlmodelc/analytics/coremldata.bin	243	d0a1721425216ae7f49e26b406963daab2c4820fe99f1cafa783361784a236d7
coreml/JointDecision.mlmodelc/coremldata.bin	632	238f58768cafeeed1937c43db581af18741f3635d0db3fe078403bf5dd1180e5
coreml/JointDecision.mlmodelc/metadata.json	2547	e01b1341757d4134b87f3cca8b613b3cc3a549590e6193a5dffad3af90d17ad7
coreml/JointDecision.mlmodelc/model.mil	7097	3c61cbc0dde3fe88083c2c664019e2a99ffb6d4fc0fbc8e921daf53ba4835b05
coreml/JointDecision.mlmodelc/weights/weight.bin	12642764	73b0288acd115fc4c36038f66b6cbf49233182488febf5d8c787bea4009535f7
coreml/Preprocessor.mlmodelc/analytics/coremldata.bin	243	22a1be9d0a96c322b682a5266c7991eaa2ee64ac4ef461a14c6d685dcc349ec6
coreml/Preprocessor.mlmodelc/coremldata.bin	634	5fba267149d02c1f3a38f8c0d56ba4dda575325850b52734921e1280bee78f26
coreml/Preprocessor.mlmodelc/metadata.json	2788	71064c5b04bc31737dd0c142dc82f4f53d64884072aaba5530675645eec90283
coreml/Preprocessor.mlmodelc/model.mil	11174	2e12d640d5c631261b9e632c458ba347153dd1dfe81a91c24b4cfc0e2c3e423b
coreml/Preprocessor.mlmodelc/weights/weight.bin	2150592	25b7de10562b2f6b51b976e1c23de4edc80cbcd6517d4d20d0205e706b4fd41a
coreml/SHA256SUMS	2202	76e4a622947ac8ef4af8cbf1bf94e1bc1463f8d67f8e9e8f450f289aded2e13f
coreml/config.json	4127	10289a3c359d2733b3f1468cfafd396da703a63925dc6d7e910dd500bc7d7982
coreml/tokens.txt	93939	d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d
onnx/SHA256SUMS	466	368204bf6425ea4e227226b83325aa2ea5a1aecda2e208f8dc3be763cb3084e8
onnx/config.json	2386	9536176118bd2fc02ffadc7f47de2ef8e98c1efd08121287a514f77feeb7bb87
onnx/decoder.onnx	47232960	3316e0084531cdbd11e2b0fe0c7a310283756ab9fec1e56f3a2b64f4f8db40c4
onnx/encoder.onnx	190236080	62d21bcce08681809ce0378ec6969fc03b7baa6c54e7d7ab161ed5b20dbba12e
onnx/joiner.onnx	25286284	a144eafbe6209bbd8f1144197d195e96f7d742611f546100776f961a31bd3513
onnx/tokens.txt	93939	d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d
onnx/vad.onnx	18209377	cf9fd348d454d16ccaf3a7474717d9054fb2be88b9b310f36854f66433d34c24