neucodec-coreml-decoder

CoreML (.mlpackage) conversion of NeuCodec decoder by Neuphonic.

Converts flat FSQ codes to 24kHz audio. Supports iOS 16+ / macOS 13+.

Files

File	Tokens	Duration
neucodec_decoder_seq50.mlpackage	50	1s
neucodec_decoder_seq100.mlpackage	100	2s
neucodec_decoder_seq250.mlpackage	250	5s

I/O

• Input codes: shape (1, 1, L) Int32, flat FSQ indices 0-65535 at 50 tok/sec
• Output audio: shape (1, 1, T) Float32 at 24kHz

Swift Usage

import CoreML

let config = MLModelConfiguration()
config.computeUnits = .cpuAndGPU
let model = try MLModel(contentsOf: compiledURL, configuration: config)

let codesArray = try MLMultiArray(shape: [1, 1, 100], dataType: .int32)
for i in 0..<100 { codesArray[i] = NSNumber(value: flatCode) }
let input = try MLDictionaryFeatureProvider(dictionary: ["codes": codesArray])
let output = try await model.prediction(from: input)
// output["audio"]: shape [1, 1, T], Float32 at 24kHz

Source

neuphonic/neucodec — Apache 2.0