moonshine-base-en-fp32-onnx

Moonshine Base English

This repository packages the FP32 ONNX artifacts for the English Moonshine Base automatic speech recognition model. The package uses an encoder-decoder pipeline with greedy cached decoding.

✨ Key Highlights

• Compact FP32 model — the encoder and decoder total 247.0 MB.
• Faster than real-time — RTFx is 22.800 on an Android™ Vivo X300 smartphone using one Arm® CPU thread.
• Low latency — median end-to-end latency is 183.705 ms for the measured 4.23-second input.
• Memory efficient — peak process RSS is 588.7 MiB.
• Arm® CPU deployment — optimized for efficient execution on Arm® CPUs.

📦 Model Details

Model Description

Moonshine Base is an encoder-decoder speech recognition model. This package provides the English FP32 ONNX encoder and merged decoder with greedy decoding and a key-value cache.

• Developed by: Moonshine AI
• Model type: English automatic speech recognition
• License: MIT, inherited from the upstream model
• Base model: moonshine-ai/moonshine
• Package form: ONNX artifacts

Model Sources

• Base model: https://huggingface.co/moonshine-ai/moonshine/tree/main/onnx/merged/base/float
• Upstream toolkit: https://github.com/moonshine-ai/moonshine

🚀 Get Started with the Model

Run with Compute Flow

Follow these steps to run Moonshine Base English on macOS® with Apple Silicon or on Linux® with an Arm®-based CPU, such as a Raspberry Pi® 5.

Step 1: Clone Compute Flow and set up the virtual environment

git clone https://github.com/ARM-software/ComputeFlow.git && cd ComputeFlow

uv sync && source .venv/bin/activate

Step 2: Install the brick and model

compute-flow update
compute-flow model install stt:moonshine moonshine-base-en
compute-flow build native

Step 3: Run the Python CLI example

python -m cli_examples.stt_cli \
  -m @manifest/moonshine/moonshine-base-en/moonshine_manifest_base_en.json \
  -a @sample/beckett.wav \
  -t 1

The example transcribes the bundled samples/beckett.wav recording. Replace -a with a path to your own WAV file.

📊 Quality evaluation

For upstream accuracy results, refer to Table 2 in the Moonshine paper, which reports Moonshine Base WER across eight OpenASR Leaderboard datasets.

🎯 Performance evaluation

Performance was measured under the following conditions:

• One Arm® CPU thread.
• One model warmup during setup, followed by 5 input-specific warmups.
• 30 consecutive measured runs with no pause between runs.
• The Android™ Vivo X300 smartphone screen was kept on.
• The model remained loaded and its state was reset between runs.

The following methodology and definitions were used:

• Runtime: ONNX Runtime 1.24.0, with KleidiAI on CPU.
• Input recording: A 4.23-second mono 16 kHz English speech recording.
• End-to-end latency is the complete speech-to-text execution time and excludes model setup.
• Average memory is the mean process RSS sampled throughout setup and inference.
• Peak memory is the maximum sampled process high-water mark.
• Latency and memory were collected in separate executions to prevent memory sampling from affecting latency.

RTFx is the total input-audio duration divided by total inference time, so higher values indicate faster processing.

⚠️ Known Limitations

This package supports English speech. It returns the complete transcription after decoding rather than streaming partial transcription text.

🛠️ Technical Specifications

Objective

Transcribe mono 16 kHz English speech into text.

Runtime Architecture

Greedy decoding does not use an external language model.

Precision and Quantization

The encoder and decoder weights and external tensor interfaces use FP32. This package is not quantized.

Input Specification

Output Specification

The model returns a complete UTF-8 transcription string.

Repository Contents

• moonshine_manifest_base_en.json — package manifest.
• encoder_model_base_en_fp32.onnx — FP32 encoder model.
• decoder_model_merged_base_en_fp32.onnx — FP32 merged decoder model.
• assets/tokenizer.bin — tokenizer asset.
• samples/beckett.wav — sample audio for a smoke test.
• metadata.yaml and benchmarks/ — model and benchmark metadata.
• SHA256SUMS — package checksums.

🗂️ Model and Asset Origin

• The encoder and decoder derive from the FP32 Moonshine Base ONNX package at moonshine-ai/moonshine.

• assets/tokenizer.bin derives from the upstream tokenizer assets.

• samples/beckett.wav derives from the Moonshine sample. It was converted to mono 16 kHz 32-bit float WAV with:

  ffmpeg -i beckett_source.wav \
      -ar 16000 \
      -ac 1 \
      -c:a pcm_f32le \
      -f wav \
      samples/beckett.wav
  

🔐 Checksums

SHA256SUMS covers every regular package file except itself and paths with a dotfile component. Verify the checked-out files from the repository root with:

shasum -a 256 -c SHA256SUMS