This repository contains Qwen3-ASR-1.7B model, which support language identification and speech recognition for 30 languages and 22 Chinese dialects, optimized for Renesas X5H platform for automatic speech recognition inference.

Model Architecture:

Model Summary:

Audio Encoder:

Parameter Value
NUM_LAYERS 24
HIDDEN_SIZE 1024
FFN_DIM 4096
NUM_HEADS 16
HEAD_DIM 64
MEL_BINS 128
OUTPUT_DIM 2048

Text Decoder:

Parameter Value
NUM_LAYERS 28
HIDDEN_SIZE 2048
FFN_DIM 6144
NUM_HEADS 16
NUM_KV_HEADS 8
HEAD_DIM 128
GROUP_SIZE 2
VOCAB_SIZE 151936
RMS_NORM_EPS 1e-6
ROPE_THETA 1000000.0
Source Model: Qwen/Qwen3-ASR-1.7B

Performance

The following performance metrics were measured with a 15-second English audio sample.

Model	Precision	Device	Offloading	Prefill Rate (tokens/sec) Batch size=1	Decode Rate (tokens/sec)
Qwen3-ASR-1.7B	W4A16	X5H - Single Cluster NPX	AudioEncode [NPX] LLM Decoder [NPX]	20.84	20.71
Qwen3-ASR-1.7B	FP16	X5H - Single Cluster NPX	AudioEncode [NPX] LLM Decoder [NPX]	12.28	12.23

Prerequisites

To run model, you need:

Renesas X5H Board with SDK v4.34.0
Hugging Face CLI: For downloading the model and installer.

Deployment

Qwen3-ASR-1.7B (W4A16)

Download the installer qwen3-asr-w4a16-runner-0.1.0-Linux.sh from Files and version tab under w4a16/binaries/rcar-x5hv1/xOS-v4.34/ folder.
Copy the installer to the X5H board and run the installer.
bash ./qwen3-asr-w4a16-runner-0.1.0-Linux.sh --prefix=./ --exclude-subdir --skip-license
Download GGUF files Qwen3-ASR-1.7B-FP16.gguf and mmproj-Qwen3-ASR-1.7b-FP16.gguf from Files and version tab under fp16 folder and copy to the installed directory on the X5H board.
Download the W4A16 coefficients from w4a16/coefficients/ and copy to the installed directory.
Expected directory structure on the X5H board.
bash
1 qwen3-asr-w4a16-runner 2 ├── Qwen3-ASR-1.7B-FP16.gguf 3 ├── mmproj-Qwen3-ASR-1.7b-FP16.gguf 4 ├── firmwares 5 ├── kernel_modules 6 ├── qwen3-asr-w4a16 7 ├── qwen3-asr-w4a16-runner 8 ├── scripts 9 ├── setup_npu.sh 10 └── test_audio 11 └── sample_en.wav

Inference - Qwen3-ASR-1.7B (W4A16)

bash

bash ./setup_npu.sh
./qwen3-asr-w4a16-runner -m Qwen3-ASR-1.7B-FP16.gguf -v mmproj-Qwen3-ASR-1.7b-FP16.gguf -w test_audio/sample_en.wav -g ./ -s

Qwen3-ASR-1.7B (FP16)

Download the installer qwen3-asr-runner-0.1.0-Linux.sh from Files and version tab under fp16/binaries/rcar-x5hv1/xOS-v4.34/ folder.
Copy the installer to the X5H board and run the installer.
bash ./qwen3-asr-runner-0.1.0-Linux.sh --prefix=./ --exclude-subdir --skip-license
Download GGUF files Qwen3-ASR-1.7B-FP16.gguf and mmproj-Qwen3-ASR-1.7b-FP16.gguf from Files and version tab under fp16 folder and copy to the installed directory on the X5H board.
Expected directory structure on the X5H board.
bash
1 qwen3-asr-runner 2 ├── Qwen3-ASR-1.7B-FP16.gguf 3 ├── mmproj-Qwen3-ASR-1.7b-FP16.gguf 4 ├── firmwares 5 ├── kernel_modules 6 ├── qwen3-asr 7 ├── qwen3-asr-runner 8 ├── scripts 9 ├── setup_npu.sh 10 └── test_audio 11 └── sample_en.wav

Inference - Qwen3-ASR-1.7B (FP16)

bash

bash ./setup_npu.sh
./qwen3-asr-runner -m Qwen3-ASR-1.7B-FP16.gguf -v mmproj-Qwen3-ASR-1.7b-FP16.gguf -w test_audio/sample_en.wav -g ./ -s

Parameter	Value
NUM_LAYERS	24
HIDDEN_SIZE	1024
FFN_DIM	4096
NUM_HEADS	16
HEAD_DIM	64
MEL_BINS	128
OUTPUT_DIM	2048

Parameter	Value
NUM_LAYERS	28
HIDDEN_SIZE	2048
FFN_DIM	6144
NUM_HEADS	16
NUM_KV_HEADS	8
HEAD_DIM	128
GROUP_SIZE	2
VOCAB_SIZE	151936
RMS_NORM_EPS	1e-6
ROPE_THETA	1000000.0

Renesas/Qwen3-ASR-1.7B-GGUF

Qwen3-ASR-1.7B - Renesas X5H

Introduction