handy-computer/parakeet-ultra-gguf

parakeet-ultra: transcribe.cpp GGUF

GGUF conversions of moondream/parakeet-ultra for use with transcribe.cpp.

Ported from upstream commit 73175eb, pinned 2026-10-04. Validated against the kestrel 0.9.1 (Photon) reference at transcribe.cpp commit 08f9e6f0 on 2026-10-05.

Moondream's post-trained parakeet-tdt-0.6b-v3. Offline multilingual speech-to-text over the same 25 European languages, with lower error rates than v3 on English and on every FLEURS language we measured. A FastConformer encoder with a TDT transducer decoder, plus a small voice-activity head that the model uses to cut long recordings at pauses into segments of at most 30 seconds. Takes 16 kHz mono WAV and produces a punctuated, cased transcript with optional token-level timestamps. A drop-in replacement for v3: same size, languages and API. Not a streaming model and does not translate.

Downloads

Quantization	Download	Size	WER (LibriSpeech test-clean)
F32	parakeet-ultra-F32.gguf	2.51 GB	1.80%
F16	parakeet-ultra-F16.gguf	1.26 GB	1.80%
Q8_0	parakeet-ultra-Q8_0.gguf	740 MB	1.80%
Q6_K	parakeet-ultra-Q6_K.gguf	611 MB	1.81%
Q5_K_M	parakeet-ultra-Q5_K_M.gguf	550 MB	1.81%
Q4_K_M	parakeet-ultra-Q4_K_M.gguf	486 MB	1.82%

WER on the full LibriSpeech test-clean split (2,620 utterances), batch size 8, timestamps none, language hint en, decoded on cuda. Measured at transcribe.cpp 021aa17d.

Greedy transducer decoding, no external LM. The reference is Moondream's own runtime (kestrel 0.9.1, the engine behind Photon), not transformers. On this manifest kestrel scores 1.80% and the C++ F32 build matches it.

Moondream's card reports 1.41% on the same split, scored with the Open ASR Leaderboard normalizer.

Long-form: on 11 full TED-LIUM 3 talks, segmented by the model's voice-activity head exactly as kestrel does it, F32 scores 2.22%, identical to kestrel, and every quant stays within 0.05pp.

Usage

Build transcribe.cpp from source:

git clone git@github.com:handy-computer/transcribe.cpp.git
cd transcribe.cpp
cmake -B build && cmake --build build

Run on a 16 kHz mono WAV:

build/bin/transcribe-cli \
  -m parakeet-ultra-Q8_0.gguf \
  input.wav

If your audio isn't already 16 kHz mono WAV, convert it first:

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

See the transcribe.cpp model page for performance numbers, numerical validation, and reproduction steps.

License

Inherited from the base model: CC-BY-4.0. See the upstream model card for full terms. This repository includes the license text (LICENSE) and a NOTICE recording how these files were derived from the original.

---

Original Model Card

The section below is reproduced from moondream/parakeet-ultra at commit 73175eb for offline reference. The upstream card is the authoritative source.

Moondream Parakeet Ultra

A post-trained version of parakeet-tdt-0.6b-v3. Same architecture, same tokenizer, same 0.6B parameters in full precision. It is better than the original on every benchmark group below, by a wide margin on the 25-language FLEURS set, in background noise and on long-form audio, and it runs faster than the original does in NeMo.

	parakeet-tdt-0.6b-v3	parakeet-ultra
Open ASR Leaderboard, 7 English sets (WER %)	6.26	5.80
FLEURS, 25 languages (WER %)	11.62	9.55
Business speech, AA-WER style (WER %)	6.15	5.79
Background noise, 9 MUSAN conditions (WER %)	6.72	5.82
TED-LIUM long-form (WER %)	2.71	1.94

Its sibling Parakeet Redux is the ternary version of the same architecture, 178 MB, built for CPUs and Apple silicon. Read the release post for the story behind both models.

Usage

Run it with Photon. Every accuracy number on this page is Photon on an NVIDIA GPU.

# pip install moondream
import moondream as md

with md.photon("moondream/parakeet-ultra") as speech:
    result = speech.transcribe(audio="speech.wav")
    print(result["text"])

    # timestamps="segment": one entry per sentence, with its start
    # and end in seconds; "word" adds the start and end of every word
    result = speech.transcribe(audio="speech.wav", timestamps="word")
    for segment in result["segments"]:
        print(segment["start"], segment["end"], segment["text"])
        for word in segment["words"]:
            print(word["start"], word["end"], word["word"])

Performance

Real-time factor: seconds of audio transcribed per second of wall clock, higher is faster. One NVIDIA B200, 128 requests in flight, the same files on both sides; NeMo 3.0 runs the original checkpoint with its own transcribe call at batch 128, Photon runs parakeet-ultra.

	NeMo, parakeet-tdt-0.6b-v3	Photon, parakeet-ultra
LibriSpeech test-clean, 2,620 utterances, 5.4 hours	6,005×	9,743×
AMI test, 12,643 utterances, 8.7 hours	4,394×	6,688×

Benchmarks

Word error rate in percent, lower is better. 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-ultra 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).

set	parakeet-tdt-0.6b-v3	parakeet-ultra
LibriSpeech test-clean	1.52	1.41
LibriSpeech test-other	3.13	2.98
AMI	10.86	9.77
Earnings-22	10.75	9.76
GigaSpeech	8.05	7.71
SPGISpeech	3.63	3.34
VoxPopuli	5.88	5.65
average	6.26	5.80

FLEURS

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

language	parakeet-tdt-0.6b-v3	parakeet-ultra
Bulgarian	11.90	10.09
Croatian	10.93	9.65
Czech	10.85	9.97
Danish	16.78	14.31
Dutch	6.18	5.46
English	4.25	3.55
Estonian	13.23	9.69
Finnish	11.05	9.19
French	4.81	4.32
German	4.13	3.61
Greek	35.71	32.25
Hungarian	13.65	10.76
Italian	2.61	2.00
Latvian	21.38	15.95
Lithuanian	21.09	16.36
Maltese	19.13	14.92
Polish	6.70	5.54
Portuguese	4.65	3.96
Romanian	11.54	9.18
Russian	5.91	5.21
Slovak	9.46	7.03
Slovene	21.76	16.64
Spanish	3.12	2.72
Swedish	13.75	11.57
Ukrainian	5.94	4.75
average	11.62	9.55

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.

set	parakeet-tdt-0.6b-v3	parakeet-ultra
AMI (cleaned)	9.52	8.48
VoxPopuli (cleaned)	3.02	3.10
Earnings-22, 30-second chunks	5.90	5.78
average	6.15	5.79

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.

set	parakeet-tdt-0.6b-v3	parakeet-ultra
LibriSpeech test-other, 10 dB	4.12	3.87
LibriSpeech test-other, 5 dB	5.49	5.10
LibriSpeech test-other, 0 dB	9.06	8.31
FLEURS German, 10 dB	5.78	4.98
FLEURS German, 5 dB	8.07	6.91
FLEURS German, 0 dB	14.45	12.08
FLEURS Spanish, 10 dB	3.99	3.13
FLEURS Spanish, 5 dB	4.12	3.38
FLEURS Spanish, 0 dB	5.44	4.64
average	6.72	5.82

Long-form

Eleven complete TED-LIUM 3 talks of 10-20 minutes each. parakeet-ultra 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.

set	parakeet-tdt-0.6b-v3	parakeet-ultra
TED-LIUM 3, 11 full talks of 10-20 minutes	2.71	1.94

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.