parakeet.md

Parakeet

NVIDIA's Parakeet family ported to transcribe.cpp. A FastConformer encoder paired with one of three decoder heads — TDT (transducer with a duration prediction head), classical RNN-T, or CTC — and a TDT+CTC hybrid that ships both heads in one checkpoint. All variants take 16 kHz mono PCM through a mel frontend; English-only across the family, except parakeet-tdt-0.6b-v3, Moondream's post-trained parakeet-ultra, Oruk's retune orukeet and the German fine-tune parakeet-primeline, which cover 25 European languages.

For the architecture deep-dive, validation contract, and porting notes, see the family doc at docs/porting/families/parakeet.md.

Choosing a variant

Most users want one of three:

• English transcription → parakeet-tdt-0.6b-v2. The default pick: small, fast, near top-of-family accuracy on English.
• Multilingual (25 European languages) → parakeet-ultra. Moondream's post-trained v3: same size, languages and API, more accurate on English and on every FLEURS language. A drop-in replacement for parakeet-tdt-0.6b-v3, which remains available. See parakeet-ultra.md for how it segments long audio.
• Multilingual, alternative → orukeet. Oruk AI's retune of v3. Same size and speed, a bit more accurate than v3, less accurate than parakeet-ultra. CC-BY-SA-4.0 weights. See orukeet.md.
• German → parakeet-primeline. primeLine's German fine-tune of v3. Same size and speed; tuned for German while keeping the other 24 v3 languages usable.
• English, alternative → phonon-2. Fermion Research's five-value-quantized repack of v3, expanded back to dense weights here. English only, accuracy on par with parakeet-tdt-0.6b-v2. CC-BY-4.0 weights. See phonon-2.md.
• Streaming / real-time → Nemotron streaming. Most Parakeet variants here are offline-only. For low-latency streaming use the FastConformer-lineage nemotron-3.5-asr-streaming-0.6b (multilingual) or nemotron-speech-streaming-en-0.6b (English). Within Parakeet itself, parakeet-unified-en-0.6b is the only streaming-capable variant.

If you specifically need the lowest WER or a different decoder:

• Lowest WER, English. parakeet-tdt-1.1b (1.38% Q8_0) is the most accurate in the family, narrowly ahead of parakeet-rnnt-1.1b (1.46%). TDT also decodes faster — the duration head lets the decoder skip frames.
• Fastest decode at any size. Use a CTC variant (parakeet-ctc-0.6b / parakeet-ctc-1.1b). Single-pass greedy alignment, no transducer loop — at a ~0.2pp WER cost vs the same-size RNN-T.
• Tiny footprint. parakeet-tdt_ctc-110m is the smallest Parakeet. The 1.1B tdt_ctc ships both heads but is primarily useful when you want TDT speed with CTC as a fallback at runtime.

All variants

WER is on LibriSpeech test-clean for the Q8_0 preset, measured by transcribe.cpp's WER pipeline. See each per-variant doc for the full quant matrix and the comparison to NVIDIA's self-reported numbers.

Variant	Params	Languages	Q8_0 size	Benchmark	Q8_0	Capabilities	Doc
parakeet-tdt-0.6b-v2	618M	en	730 MB	LibriSpeech test-clean (WER)	1.69%	token timestamps	parakeet-tdt-0.6b-v2.md
parakeet-ultra	627M	25 languages + auto-detect	740 MB	LibriSpeech test-clean (WER)	1.80%	token timestamps	parakeet-ultra.md
parakeet-tdt-0.6b-v3	627M	25 languages + auto-detect	740 MB	LibriSpeech test-clean (WER)	1.94%	token timestamps	parakeet-tdt-0.6b-v3.md
orukeet	627M	25 languages + auto-detect	740 MB	LibriSpeech test-clean (WER)	1.86%	token timestamps	orukeet.md
phonon-2	627M	en	740 MB	LibriSpeech test-clean (WER)	2.15%	token timestamps	phonon-2.md
parakeet-primeline	627M	25 languages + auto-detect	740 MB	FLEURS de (WER)	5.98%	token timestamps	parakeet-primeline.md
parakeet-tdt-1.1b	1.1B	en	1.27 GB	LibriSpeech test-clean (WER)	1.38%	token timestamps	parakeet-tdt-1.1b.md
parakeet-tdt_ctc-110m	114M	en	135 MB	LibriSpeech test-clean (WER)	2.43%	token timestamps	parakeet-tdt_ctc-110m.md
parakeet-tdt_ctc-1.1b	1.1B	en	1.27 GB	LibriSpeech test-clean (WER)	1.87%	token timestamps	parakeet-tdt_ctc-1.1b.md
parakeet-rnnt-0.6b	617M	en	730 MB	LibriSpeech test-clean (WER)	1.62%	token timestamps	parakeet-rnnt-0.6b.md
parakeet-rnnt-1.1b	1.1B	en	1.27 GB	LibriSpeech test-clean (WER)	1.46%	token timestamps	parakeet-rnnt-1.1b.md
parakeet-ctc-0.6b	609M	en	722 MB	LibriSpeech test-clean (WER)	1.87%	token timestamps	parakeet-ctc-0.6b.md
parakeet-ctc-1.1b	1.1B	en	1.26 GB	LibriSpeech test-clean (WER)	1.85%	token timestamps	parakeet-ctc-1.1b.md
parakeet-unified-en-0.6b	618M	en	731 MB	LibriSpeech test-clean (WER)	1.60%	streaming, token timestamps	parakeet-unified-en-0.6b.md

* parakeet-primeline is scored on FLEURS German (862 utterances), not LibriSpeech test-clean, so its number is not comparable to the rest of the column. Its NeMo reference on the same manifest is 5.98%.

Pre-built GGUFs for every variant and quant are hosted under handy-computer on Hugging Face; each per-variant doc has direct download links.

Input limits

No practical per-call length limit (transcribe_capabilities.max_audio_ms == 0): the Conformer encoder's positional encoding is recomputed per call, so audio of any length is processed in a single pass — pass arbitrarily long recordings. The exception is parakeet-ultra, which cuts audio over 30 s into segments at pauses found by its voice-activity head; length is still unbounded. See the input-length contract.

Long-form audio

Variants whose checkpoint uses full self-attention (parakeet-tdt-0.6b-v2, parakeet-tdt-0.6b-v3, orukeet, parakeet-primeline, the 1.1B and 110M TDT/RNN-T/CTC variants) run NeMo's rel_pos_local_attn window of [256, 256] encoder frames (20.5 s each side). NVIDIA documents this window as the family's long-audio mode (change_attention_model("rel_pos_local_attn", [256, 256])); NeMo itself defaults to full attention, which it rates for up to 24 minutes on an 80 GB A100. Clips up to 20.5 s run the full-attention graph, so their output is identical to full attention; beyond that, encoder time and memory grow linearly with length instead of quadratically, so a 25-minute talk runs on a 16 GB laptop and multi-hour recordings fit. parakeet-tdt_ctc-1.1b ships its own [128, 128] window and parakeet-ultra keeps full attention inside its 30 s segments; the streaming variants keep their chunked attention.

The port matches NeMo with the same window at tensor level (encoder blocks and enc.final within the family tolerances; six variants on a 70 s clip, the bounded-window graph, and v3 on a 35 s clip, the dense local graph), so the accuracy below is the model's rather than the port's.

WER, full attention -> window (Q8_0):

set	v3	v2	tdt-1.1b	orukeet	primeline
TED-LIUM 3 full talks (10-25 min)	2.87 -> 2.71	3.20 -> 3.11	4.98 -> 3.94	2.94 -> 3.06	3.57 -> 3.82
TED-LIUM 3 talks cut into 1-5 min pieces	3.19 -> 2.35	3.06 -> 3.09	4.09 -> 3.91	3.37 -> 3.01	3.89 -> 4.00
LibriSpeech test-clean, 1-5 min clips	2.52 -> 2.16	1.63 -> 1.74	1.31 -> 1.26	1.90 -> 1.86	2.69 -> 2.65

On the other LibriSpeech-clip variants the window is equal or better (rnnt-0.6b 1.62 -> 1.51, rnnt-1.1b 1.31 -> 1.31, ctc-0.6b 1.86 -> 1.79, ctc-1.1b 2.11 -> 1.95, tdt_ctc-110m 2.51 -> 2.42).

Non-English (Multilingual TEDx test talks): on 1-5 min clips the window is equal or better for orukeet in every language except Portuguese (10.1 -> 10.8), and for parakeet-tdt-0.6b-v3 except French (15.5 -> 16.4) and Spanish; parakeet-primeline loses on French (10.7 -> 11.8), Portuguese and Spanish. On full talks v3's French drops from 12.4% to 17.7%. parakeet-tdt-0.6b-v3 and its fine-tunes can drift into English on accented spontaneous speech in either mode; under full attention that drift can take over a whole talk, under the window it is confined to the window.

Quick start

Pick a variant and run:

cmake -B build
cmake --build build

build/bin/transcribe-cli \
  -m models/parakeet-tdt-0.6b-v2/parakeet-tdt-0.6b-v2-Q8_0.gguf \
  samples/jfk.wav

The repo doesn't ship the GGUFs — pull them from the corresponding handy-computer/<variant>-gguf repo on Hugging Face, or convert from the upstream NVIDIA .nemo checkpoint via the per-variant doc's reproduction section.

Capabilities

All Parakeet variants support:

• Transcription of 16 kHz mono WAV input.
• Token-level timestamps at the encoder frame rate (TDT and RNN-T; CTC also exposes frame-level alignment).

Buffered streaming is supported on parakeet-unified-en-0.6b across all six published (L, C, R) configurations from the model's training menu (lookahead latency from 160ms at (70, 1, 1) through 2.08s at the default (70, 13, 13)). See parakeet-unified-en-0.6b.md for the per-config WER and the --stream-buf-{left,chunk,right}-ms CLI surface. Other Parakeet variants run offline only.

What's not supported (consistent across the family): translation, speaker diarization, and VAD as a user-facing feature (parakeet-ultra uses its voice-activity head internally, only to segment long audio). Language coverage is English-only except parakeet-tdt-0.6b-v3, parakeet-ultra and parakeet-primeline (25 European languages, auto-detected; the language hint is accepted but does not change the output). Note that the v3 lineage, including parakeet-primeline, writes German ss where standard orthography uses ß; see parakeet-primeline.md for why and what to do about it. See the family doc for the full runtime contract.