Automatic Speech Recognition - ASR

This example runs a speech recognition model based on the hbm_runtime inference engine to automatically transcribe .wav audio files into corresponding text. The example code is located in the /app/pydev_demo/07_speech_sample/01_asr/ directory.

Model Description

• Introduction:

  The ASR (Automatic Speech Recognition) model converts audio signals into text. The input is a single-channel audio waveform (after sample rate conversion and normalization), and the output is a character-level token sequence. When used together with a vocabulary (vocab) file, it supports Chinese speech transcription. This example uses a quantized .hbm model.

• HBM Model Name: asr.hbm

• Input Format: Audio waveform, single-channel, sampled at 16kHz, with a maximum length of 30,000 samples.

• Output: Probability distribution (logits) over character tokens. After decoding via argmax, the logits are mapped to the recognized text.

• Model Download URL (automatically downloaded by the program):

  https://archive.d-robotics.cc/downloads/rdk_model_zoo/rdk_s100/asr/asr.hbm

Functionality Description

• Model Loading

  Load the ASR model using hbm_runtime, which automatically parses the model's input/output shapes and quantization information.

• Input Preprocessing

  Use SoundFile to read audio files (supports .wav), and process the audio as follows:

  • Convert to single-channel
  • Resample to the target sample rate (default: 16kHz)
  • Normalize to zero-mean and unit-variance (z-score)
  • Pad or truncate to a fixed length (e.g., 30,000 samples)
  • Support generator-based processing for long audio inputs, enabling streaming recognition.

• Inference Execution

  Perform inference using the .run() method, producing a logits tensor as output.

• Post-processing

  Use np.argmax() to obtain token indices from the output logits, then map them to characters using the vocab dictionary file (in JSON format) to produce the final transcribed text.

Environment Dependencies

• Ensure that the dependencies in pydev are installed:

  pip install -r ../../requirements.txt

• Install the soundfile package:

  pip install soundfile==0.13.1

Directory Structure

01_asr/
├── asr.py                      # Main inference script

Parameter Description

Parameter	Description	Default Value
--model-path	Path to the model file (.hbm format)	/opt/hobot/model/s100/basic/asr.hbm
--audio-file	Input audio file (supports .wav or .flac)	/app/res/assets/chi_sound.wav
--vocab-file	Vocabulary file mapping tokens to IDs	/app/res/labels/vocab.json
--priority	Inference priority (0–255; higher values indicate higher priority)	0
--bpu-cores	Specify which BPU cores to use (e.g., --bpu-cores 0 1)	[0]
--audio_maxlen	Fixed length after audio padding/truncation (in samples)	30000
--new_rate	Target sample rate; audio will be automatically resampled	16000

Quick Start

• Run the model

  • Using default parameters:

    python asr.py

  • Running with specified parameters:

    python asr.py \
    --model-path /opt/hobot/model/s100/basic/asr.hbm \
    --audio-file /app/res/assets/chi_sound.wav \
    --vocab-file /app/res/labels/vocab.json \
    --priority 0 \
    --bpu-cores 0 \
    --audio_maxlen 30000 \
    --new_rate 16000

• View Results

  Upon successful execution, the result will be printed:

  I am Qwen, a large-scale language model developed by Alibaba Cloud.

Notes

• If the specified model path does not exist, the program will attempt to download the model automatically.