ASR (Automatic Speech Recognition)#

Complete workflow for speech recognition with audio understanding capabilities.

Example model: Qwen3-ASR-0.6B

Prerequisites: Complete the Installation Guide before proceeding.

Step 1: Export (x86 Host)#

export WORKSPACE_DIR=$HOME/tensorrt-edgellm-workspace
export MODEL_NAME=Qwen3-ASR-0.6B
mkdir -p $WORKSPACE_DIR
cd $WORKSPACE_DIR

# Export language model and audio encoder
trtedge export \
  Qwen/Qwen3-ASR-0.6B \
  $MODEL_NAME/onnx

Step 2: Transfer to Device#

# Transfer ONNX to device
scp -r $MODEL_NAME/onnx \
  <device_user>@<device_ip>:~/tensorrt-edgellm-workspace/$MODEL_NAME/

Step 3: Build Engines (Thor Device)#

# Set up workspace directory on device
export WORKSPACE_DIR=$HOME/tensorrt-edgellm-workspace
export MODEL_NAME=Qwen3-ASR-0.6B
cd /path/to/TensorRT-Edge-LLM

# Build language model engine
./build/examples/llm/llm_build \
  --onnx-dir $WORKSPACE_DIR/$MODEL_NAME/onnx/llm \
  --engine-dir $WORKSPACE_DIR/$MODEL_NAME/engines/llm \
  --max-batch-size 1 \
  --max-input-len 1024 \
  --max-kv-cache-capacity 4096

# Build audio encoder engine
./build/examples/multimodal/audio_build \
  --onnx-dir $WORKSPACE_DIR/$MODEL_NAME/onnx/audio \
  --engine-dir $WORKSPACE_DIR/$MODEL_NAME/engines/audio \
  --min-time-steps 100 \
  --max-time-steps 3000

Step 4: Run Inference (Thor Device)#

Audio decode + mel-spectrogram extraction run in the C++ runtime via vendored miniaudio + an in-tree mel extractor; no offline Python preprocessing step is required. Supported containers: .wav, .mp3, .flac. The mel pipeline (whisper / parakeet) is auto-derived from the engine’s audio/config.json, pinned by the model — mirroring HF / vLLM where the FE choice ships with the model checkpoint.

Create an input file $WORKSPACE_DIR/input_asr.json (replace /path/to/audio.wav with the actual audio file path):

{
    "batch_size": 1,
    "temperature": 1.0,
    "top_p": 1.0,
    "top_k": 50,
    "max_generate_length": 256,
    "requests": [
        {
            "messages": [
                {
                    "role": "system",
                    "content": ""
                },
                {
                    "role": "user",
                    "content": [{"type": "audio", "audio": "/path/to/audio.wav"}]
                }
            ]
        }
    ]
}

Run inference:

cd /path/to/TensorRT-Edge-LLM

./build/examples/llm/llm_inference \
  --engine-dir $WORKSPACE_DIR/$MODEL_NAME/engines/llm \
  --multimodal-engine-dir $WORKSPACE_DIR/$MODEL_NAME/engines/audio \
  --input-file $WORKSPACE_DIR/input_asr.json \
  --output-file $WORKSPACE_DIR/output_asr.json

Check output_asr.json for the speech recognition transcription.

Quantization (Optional)#

For deployments where engine size matters more than peak accuracy, the new pipeline supports independent precision selection for the LLM backbone and the audio encoder. Numbers below are measured on Qwen3-ASR-0.6B; 1.7B is noted where it diverges.

Supported precision combinations#

LLM backbone

Audio encoder

0.6B status

1.7B status

FP16

FP16

✅

✅

FP8

FP16

✅

✅

FP8

FP8

✅

✅

NVFP4

FP16

✅

✅

NVFP4

FP8

❌ Empty output (combined quant noise exceeds the first-token EOS-vs-correct logit margin on 0.6B)

✅ (1.7B tolerates the combined noise)

Mixed-precision rules (tensorrt_edgellm/scripts/quantize.py):

  • --quantization quantizes the LLM backbone only.

  • --audio-quantization fp8 opts the audio tower in. Omit it to keep the audio tower at FP16 regardless of --quantization – audio encoders are quantization-sensitive, so the default is conservative (mirrors the --visual-quantization behaviour for VLMs).

  • --lm-head-quantization is similarly opt-in.