Moondream Parakeet Redux

A 1.58-bit version of parakeet-tdt-0.6b-v3. Same architecture, same tokenizer, but every encoder weight is -1, 0 or +1. It fits in 178 MB, runs at 113× real time on eight x86 CPU cores, 2.5× the fastest other Parakeet runtime we measured, and stays within 0.3 WER of the original on English while beating it on the 25-language FLEURS set and on long-form audio.

Its sibling Parakeet Ultra is the full-precision version of the same architecture, trained further, for GPUs: better than the original on every benchmark. Read the release post for the story behind both models.

Usage

Run it with Photon, which reads the packed weights directly: AVX-512 VNNI on x86, NEON on ARM, Metal on Apple GPUs. Every speed number on this page is Photon.

# pip install moondream  (2.4.0 or later)
import moondream as md

# device: "cpu", "mps" or "cuda"; leave it out to take CUDA,
# then Apple silicon, then the CPU
with md.photon("moondream/parakeet-redux", device="cpu") as speech:
    result = speech.transcribe(audio="speech.wav")
    print(result["text"])
    # From the undersurface of the clouds there are continual
    # emissions of lurid light. Electric matter is in continual
    # evolution from their component molecules. The gaseous ...

    # timestamps="segment": one entry per sentence, with its start
    # and end in seconds
    result = speech.transcribe(audio="speech.wav", timestamps="segment")
    for segment in result["segments"]:
        print(segment["start"], segment["end"], segment["text"])
    # 0.32 5.12 From the undersurface of the clouds there are ...
    # 5.12 9.92 Electric matter is in continual evolution from ...
    # 9.92 21.04 The gaseous elements of the air need to be ...

    # timestamps="word": the same sentences, each with the start
    # and end of every word
    result = speech.transcribe(audio="speech.wav", timestamps="word")
    for word in result["segments"][0]["words"][:3]:
        print(word["start"], word["end"], word["word"])
    # 0.32 0.56 From
    # 0.56 0.72 the
    # 0.72 1.44 undersurface

Performance

Real-time factor: seconds of audio transcribed per second of wall clock, one utterance at a time, higher is faster. The other rows are the fastest other ways we found to run Parakeet on the same machine and the same audio.

How this was measured

One utterance at a time. Incumbents: parakeet.cpp (ggml), sherpa-onnx and onnx-asr (ONNX Runtime) and, on the Mac, parakeet-mlx, each at its own defaults. x86: LibriSpeech test-clean (2,620 utterances), every runtime on the same 8 cores. Apple silicon: a 50-utterance slice of LibriSpeech dev-clean, with cool-downs between runs. WER is scored the same way as the benchmarks below.

x86 CPU

AMD EPYC 9575F (Zen 5, up to 5.0 GHz, AVX-512), 8 physical cores of one chiplet, DDR5-6000, Ubuntu 22.04.

Apple silicon

MacBook Air with the Apple M2 (4 performance and 4 efficiency CPU cores, 10-core GPU), 16 GB unified memory, macOS 15.

Benchmarks

Word error rate in percent, lower is better. parakeet-redux is better than the original on FLEURS and on long-form audio, close to it on English, and behind it in noise. Both models are scored on the same files with the Open ASR Leaderboard's own pipeline as of September 2026 (its normalizers and compound-merging alignment, with the FLEURS references prepared as the leaderboard's text column is). parakeet-redux runs in Photon on an NVIDIA GPU, the original in NeMo in bf16.

Open ASR Leaderboard

The seven English test sets of the Hugging Face Open ASR Leaderboard: audiobooks (LibriSpeech), meetings (AMI), earnings calls (Earnings-22), podcasts and YouTube (GigaSpeech), financial calls (SPGISpeech) and parliament (VoxPopuli).

FLEURS

The FLEURS test split: read Wikipedia sentences, a few hundred per language, for all 25 languages the model supports.

Business speech (AA-WER style)

The AMI and VoxPopuli test sets with the Artificial Analysis cleaning applied, and Earnings-22 scored in 30-second chunks joined per call, the way the AA-WER benchmark does it.

Background noise

The clean sets with MUSAN noise (the half of the corpus not used for training) mixed in at a fixed signal-to-noise ratio; 0 dB means the noise is as loud as the speech.

Noise is where the gap to the original is widest: the ternary encoder's acoustic margin is thinner, and at low SNR it substitutes similar-sounding words more often. Dropped or invented content is not more frequent than the original's.

Long-form

Eleven complete TED-LIUM 3 talks of 10-20 minutes each. parakeet-redux transcribes them through Photon, whose segmenter cuts each talk at pauses found by the model's VAD head into segments of at most 30 seconds; the original runs NeMo's own long-audio path.

Notes

• Based on parakeet-tdt-0.6b-v3 by NVIDIA. Languages, tokenizer and output conventions (punctuation, casing, numerals) are the original's.
• Runs with Photon.
• Long audio is segmented by the model itself: the weights carry a small voice-activity head on the encoder's subsampler, and Photon uses it to cut recordings at pauses into segments of at most 30 seconds. No external VAD model is needed.
• License is CC-BY-4.0, same as the original.