Modelpublic

aoiandroid/nemotron-3.5-asr-streaming-0.6b-coreml

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes19downloads

Model Card

Nemotron-3.5-ASR-Streaming-0.6B Multilingual (Core ML FP16)

This repository contains the official Apple Core ML (.mlmodel) export of NVIDIA's Nemotron 3.5 ASR Streaming 0.6B Multilingual model (nvidia/nemotron-3.5-asr-streaming-0.6b).

Converted with coremltools and Float16 weight quantization, optimized for real-time streaming Automatic Speech Recognition (ASR) in 50+ languages (including English, Japanese, Spanish, French, German, Korean, Chinese, etc.) on Apple Neural Engine (ANE), Apple Silicon GPU, and CPU.

📦 Model Artifacts

Model File	Description	Inputs	Outputs	Precision	Size
`Nemotron_FastConformer_Encoder_Part1.mlmodel`	Subsampling Conv + FastConformer Layers 0–11	`input_features` `[1, T, 128]`	`part1_hidden_states` `[1, T_sub, 1024]`, `position_embeddings` `[1, T_pos, 1024]`	FP16	~290 MB
`Nemotron_FastConformer_Encoder_Part2.mlmodel`	FastConformer Layers 12–23 + Encoder Projector (1024 → 640)	`part1_hidden_states` `[1, T_sub, 1024]`, `position_embeddings` `[1, T_pos, 1024]`	`encoder_projected_states` `[1, T_sub, 640]`	FP16	~290 MB
`Nemotron_RNNT_Decoder_Joint.mlmodel`	Predictor / RNN-T Decoder + 4D Joint Network	`decoder_input_ids` `[1, U]`, `encoder_hidden_states` `[1, T_sub, 640]`	`joint_logits` `[1, T_sub, U, 13088]`	FP16	~25 MB

🌐 Supported Languages (50+ Languages)

en (English), ja (Japanese), es (Spanish), de (German), fr (French), it (Italian), ar (Arabic), ko (Korean), pt (Portuguese), ru (Russian), hi (Hindi), zh (Chinese), vi (Vietnamese), he (Hebrew), nl (Dutch), cs (Czech), da (Danish), pl (Polish), no (Norwegian), sv (Swedish), th (Thai), tr (Turkish), bg (Bulgarian), el (Greek), et (Estonian), fi (Finnish), hr (Croatian), hu (Hungarian), lt (Lithuanian), lv (Latvian), ro (Romanian), sk (Slovak), uk (Ukrainian), mt (Maltese), sl (Slovenian), etc.

🚀 Swift Usage Example (iOS 16+ / macOS 13+)

swift

import CoreML

// Load Core ML Models
let encoderP1 = try Nemotron_FastConformer_Encoder_Part1(configuration: MLModelConfiguration())
let encoderP2 = try Nemotron_FastConformer_Encoder_Part2(configuration: MLModelConfiguration())
let decoderJoint = try Nemotron_RNNT_Decoder_Joint(configuration: MLModelConfiguration())

// 1. Run Encoder Part 1 on 128-bin Mel Features
let p1Output = try encoderP1.prediction(input_features: melSpectrogramMLArray)

// 2. Run Encoder Part 2 to produce 640-dim Projected States
let p2Output = try encoderP2.prediction(
    part1_hidden_states: p1Output.part1_hidden_states,
    position_embeddings: p1Output.position_embeddings
)

// 3. RNN-T Greedy Search Loop
let jointOutput = try decoderJoint.prediction(
    decoder_input_ids: currentTokenIDs,
    encoder_hidden_states: p2Output.encoder_projected_states
)

🛠 Conversion Details

—Base PyTorch Model: nvidia/nemotron-3.5-asr-streaming-0.6b
—Precision: Float16 weight quantization (nbits=16)
—Conversion Tool: coremltools 8.3
—Target OS: iOS 16+ / macOS 13+ / tvOS 16+