cjpais · Oct 5, 2026 · Oct 5, 2026 · Oct 5, 2026 · Oct 5, 2026 · Oct 5, 2026
@@ -23,7 +23,7 @@ C/C++ speech-to-text inference library. Runs diverse STT model families via [GGU
 | Multitalker Parakeet Streaming 0.6B v1 | `multitalker-parakeet-streaming-0.6b-v1` | diarize, streaming, token timestamps | [docs/models/multitalker-parakeet-streaming-0.6b-v1.md](docs/models/multitalker-parakeet-streaming-0.6b-v1.md) |
 | Nemotron 3.5 ASR Streaming 0.6B | `nemotron-3.5-asr-streaming-0.6b` | streaming, token timestamps | [docs/models/nemotron-3.5-asr-streaming-0.6b.md](docs/models/nemotron-3.5-asr-streaming-0.6b.md) |
 | Nemotron Speech Streaming EN 0.6B | `nemotron-speech-streaming-en-0.6b` | streaming, token timestamps | [docs/models/nemotron-speech-streaming-en-0.6b.md](docs/models/nemotron-speech-streaming-en-0.6b.md) |
-| Parakeet | `parakeet-ctc-0.6b`, `parakeet-ctc-1.1b`, `parakeet-primeline`, `parakeet-rnnt-0.6b`, `parakeet-rnnt-1.1b`, `parakeet-tdt-0.6b-v2`, `parakeet-tdt-0.6b-v3`, `parakeet-tdt-1.1b`, `parakeet-tdt_ctc-1.1b`, `parakeet-tdt_ctc-110m`, `parakeet-unified-en-0.6b` | streaming, token timestamps | [docs/models/parakeet.md](docs/models/parakeet.md) |
+| Parakeet | `parakeet-ctc-0.6b`, `parakeet-ctc-1.1b`, `parakeet-primeline`, `parakeet-rnnt-0.6b`, `parakeet-rnnt-1.1b`, `parakeet-tdt-0.6b-v2`, `parakeet-tdt-0.6b-v3`, `parakeet-tdt-1.1b`, `parakeet-tdt_ctc-1.1b`, `parakeet-tdt_ctc-110m`, `parakeet-ultra`, `parakeet-unified-en-0.6b` | streaming, token timestamps | [docs/models/parakeet.md](docs/models/parakeet.md) |
 | Qwen3-ASR | `qwen3-asr-0.6b`, `qwen3-asr-1.7b` | - | [docs/models/qwen3-asr.md](docs/models/qwen3-asr.md) |
 | SenseVoice Small | `sensevoice-small` | - | [docs/models/sensevoice-small.md](docs/models/sensevoice-small.md) |
 | Voxtral (2507) | `voxtral-mini-3b-2507`, `voxtral-small-24b-2507` | translate | [docs/models/voxtral.md](docs/models/voxtral.md) |

@@ -0,0 +1,251 @@
+# Parakeet Ultra
+
+<!-- catalog:intro -->
+Upstream: [`moondream/parakeet-ultra`](https://huggingface.co/moondream/parakeet-ultra) at [`73175eb`](https://huggingface.co/moondream/parakeet-ultra/commit/73175eb).
+
+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.
+<!-- /catalog -->
+
+## What it's for
+
+Offline multilingual speech-to-text, and the default pick over
+`parakeet-tdt-0.6b-v3` for new work. Moondream post-trained v3 without
+changing its architecture, tokenizer or frontend, so it covers the same 25
+European languages and drops into any v3 setup by swapping the GGUF.
+
+The one addition is a small voice-activity head on the encoder's input
+layers. The model uses it to segment long audio, described below. It is not
+a streaming model and does not translate.
+
+See Moondream's [model card](https://huggingface.co/moondream/parakeet-ultra)
+for training details and their own evaluation.
+
+## Long-form audio
+
+Audio up to 30 seconds is transcribed in one pass, exactly like v3. Longer
+audio is segmented the way Moondream's runtime does it:
+
+1. The voice-activity head scores every 80 ms frame, scanning the file in
+   120 s blocks.
+2. Frames above 0.5 are speech; gaps under 0.1 s are bridged and runs under
+   0.1 s dropped. A pause is any silence of at least 0.2 s.
+3. Each segment ends at the midpoint of the last pause that falls between
+   1 s and 30 s into it, or at 30 s if there is none. Segments without speech
+   are skipped.
+4. Each segment is transcribed independently; the texts are joined with a
+   space and timestamps are shifted to file time.
+
+The thresholds are stored in the GGUF, not hard-coded. Cut points match
+Moondream's runtime sample for sample. Two consequences when moving from v3:
+segment boundaries and timestamps on long files differ from v3's output,
+and audio over 30 s pays for one extra voice-activity pass.
+
+<!-- catalog:pin -->
+Licensed CC-BY-4.0. Ported from upstream commit [`73175eb`](https://huggingface.co/moondream/parakeet-ultra/commit/73175eb), pinned 2026-10-04. Validated against the kestrel 0.9.1 (Photon) reference at transcribe.cpp commit [`08f9e6f0`](https://github.com/handy-computer/transcribe.cpp/tree/08f9e6f0) on 2026-10-05.
+<!-- /catalog -->
+
+## Download
+
+<!-- catalog:downloads -->
+| Quantization | Download |    Size | WER (LibriSpeech test-clean) |
+| --- | --- | ---: | ---: |
+| F32          | [parakeet-ultra-F32.gguf](https://huggingface.co/handy-computer/parakeet-ultra-gguf/resolve/main/parakeet-ultra-F32.gguf) | 2.51 GB | 1.80% |
+| F16          | [parakeet-ultra-F16.gguf](https://huggingface.co/handy-computer/parakeet-ultra-gguf/resolve/main/parakeet-ultra-F16.gguf) | 1.26 GB | 1.80% |
+| Q8_0         | [parakeet-ultra-Q8_0.gguf](https://huggingface.co/handy-computer/parakeet-ultra-gguf/resolve/main/parakeet-ultra-Q8_0.gguf) |  740 MB | 1.80% |
+| Q6_K         | [parakeet-ultra-Q6_K.gguf](https://huggingface.co/handy-computer/parakeet-ultra-gguf/resolve/main/parakeet-ultra-Q6_K.gguf) |  611 MB | 1.81% |
+| Q5_K_M       | [parakeet-ultra-Q5_K_M.gguf](https://huggingface.co/handy-computer/parakeet-ultra-gguf/resolve/main/parakeet-ultra-Q5_K_M.gguf) |  550 MB | 1.81% |
+| Q4_K_M       | [parakeet-ultra-Q4_K_M.gguf](https://huggingface.co/handy-computer/parakeet-ultra-gguf/resolve/main/parakeet-ultra-Q4_K_M.gguf) |  486 MB | 1.82% |
+<!-- /catalog -->
+
+<!-- catalog:recipe -->
+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`.
+<!-- /catalog -->
+
+<!-- catalog:prose field=wer.notes -->
+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.
+<!-- /catalog -->
+
+<!-- catalog:accuracy -->
+**FLEURS test**
+
+| Language | Metric |   Q8_0 |
+| --- | --- | ---: |
+| bg       | WER    | 11.04% |
+| cs       | WER    | 11.04% |
+| da       | WER    | 15.98% |
+| de       | WER    |  4.58% |
+| el       | WER    | 32.25% |
+| en       | WER    |  4.26% |
+| es       | WER    |  3.06% |
+| et       | WER    | 14.44% |
+| fi       | WER    | 11.14% |
+| fr       | WER    |  4.66% |
+| hr       | WER    | 11.30% |
+| hu       | WER    | 13.09% |
+| it       | WER    |  2.34% |
+| lt       | WER    | 17.06% |
+| lv       | WER    | 18.04% |
+| mt       | WER    | 17.01% |
+| nl       | WER    |  6.70% |
+| pl       | WER    |  6.16% |
+| pt       | WER    |  4.16% |
+| ro       | WER    | 10.12% |
+| ru       | WER    |  5.63% |
+| sk       | WER    |  7.83% |
+| sl       | WER    | 18.98% |
+| sv       | WER    | 12.83% |
+| uk       | WER    |  5.71% |
+<!-- /catalog -->
+
+The voice-activity head is kept at F32 in every quant. Lower quants can still
+move a cut point, because the head reads the quantized input layers' output:
+on a 306 s test file, F16, Q8_0 and Q6_K cut exactly where F32 does, Q5_K_M
+differs at one segment boundary and Q4_K_M produces 12 segments instead of
+11.
+
+## Quick Start
+
+```bash
+cmake -B build
+cmake --build build
+
+build/bin/transcribe-cli \
+  -m models/parakeet-ultra/parakeet-ultra-Q8_0.gguf \
+  samples/jfk.wav
+```
+
+If your audio is not already 16 kHz mono WAV, convert it first:
+
+```bash
+ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav
+```
+
+## Performance
+
+### Apple M4 Max
+
+<!-- catalog:perf machine=m4-max -->
+Compute latency (mel + encode + decode), speedup over realtime in parentheses; profile `asr-publication-v2`: mean over 3 iterations after 1 warmup.
+
+| Backend | Sample       |             Q8_0 |           Q4_K_M |
+| ------- | ------------ | ---------------: | ---------------: |
+| Metal   | jfk (11.0s)  |  53 ms (206.94×) |  54 ms (202.43×) |
+| Metal   | dots (35.3s) | 206 ms (171.31×) | 212 ms (166.34×) |
+| CPU     | jfk (11.0s)  |  283 ms (38.87×) |  311 ms (35.42×) |
+| CPU     | dots (35.3s) |  1.13 s (31.20×) |  1.12 s (31.68×) |
+
+Apple M4 Max: transcribe.cpp `a91a58db` on 2026-10-05.
+<!-- /catalog -->
+
+### AMD Ryzen 7 PRO 4750U
+
+<!-- catalog:perf machine=ryzen-4750u -->
+Compute latency (mel + encode + decode), speedup over realtime in parentheses; profile `asr-publication-v2`: mean over 3 iterations after 1 warmup.
+
+| Backend | Sample       |            Q8_0 |          Q4_K_M |
+| ------- | ------------ | --------------: | --------------: |
+| Vulkan  | jfk (11.0s)  | 455 ms (24.18×) | 459 ms (23.98×) |
+| Vulkan  | dots (35.3s) | 1.62 s (21.82×) | 1.64 s (21.55×) |
+| CPU     | jfk (11.0s)  | 733 ms (15.01×) | 778 ms (14.13×) |
+| CPU     | dots (35.3s) | 3.05 s (11.57×) | 3.18 s (11.12×) |
+
+AMD Ryzen 7 PRO 4750U (Radeon RADV RENOIR): transcribe.cpp `08f90613` on 2026-10-05.
+<!-- /catalog -->
+
+Benchmark reproduction:
+
+```bash
+uv run scripts/bench/run.py --profile --models parakeet-ultra
+```
+
+## Numerical Validation
+
+transcribe.cpp is validated tensor-by-tensor against kestrel 0.9.1, the
+runtime Moondream ships for this model, on three cases: `jfk` (encoder and
+decode), and `dots` (35 s) and `dots-full` (306 s) through the long-form path.
+The long-form cases compare every voice-activity scan block, the cut points
+(exact), each segment's encoder output and the stitched transcript. All
+transcripts match kestrel verbatim, and LibriSpeech test-clean and TED-LIUM
+long-form WER at F32 equal kestrel's exactly.
+
+| Field | Value |
+| --- | --- |
+| Reference | kestrel 0.9.1 / kestrel-kernels 0.7.4, `moondream/parakeet-ultra` |
+| Dump script | `scripts/dump_reference_parakeet_kestrel.py` |
+| Manifest | `tests/golden/parakeet/parakeet-ultra.manifest.json` |
+| Tolerances | `tests/tolerances/parakeet-ultra.json` |
+| Command | `uv run scripts/validate.py all --family parakeet --variant parakeet-ultra` |
+
+The tensor comparison runs with `TRANSCRIBE_NO_FLASH=1` (set by the manifest),
+because the flash-attention path rounds one attention term to F16. Flash
+attention stays on in production; WER is the check there. Remaining drift is
+float32 accumulation order, plus float32 STFT noise on kestrel's side of the
+mel comparison.
+
+To match kestrel, the model takes three runtime paths that the other parakeet
+variants don't, each switched on by a GGUF key: masking after every strided
+subsampling conv, F32 for the subsampler's pointwise convs, and a single
+per-utterance symbol budget in the TDT decoder.
+
+## Reproduction
+
+### Convert
+
+Moondream ships HF safetensors only; the converter maps them onto the NeMo
+layout and adds the voice-activity head and segmenter constants.
+
+```bash
+uv run --project scripts/envs/parakeet \
+  scripts/convert-parakeet.py moondream/parakeet-ultra
+```
+
+### Quantize
+
+```bash
+uv run scripts/quantize-all.py \
+  models/parakeet-ultra/parakeet-ultra-F32.gguf
+```
+
+### Validate
+
+```bash
+uv run scripts/validate.py all --family parakeet --variant parakeet-ultra
+```
+
+### WER
+
+```bash
+uv run scripts/wer/ingest.py librispeech
+uv run scripts/wer/ingest.py tedlium-longform
+
+# reference arm (kestrel)
+uv run --project scripts/envs/parakeet-kestrel \
+  scripts/wer/run_reference_parakeet_kestrel.py \
+    --manifest samples/wer/tedlium-longform.manifest.jsonl \
+    --out reports/wer/parakeet-ultra-REF.tedlium-longform.jsonl
+
+# transcribe.cpp arm
+uv run scripts/wer/run.py \
+  --model models/parakeet-ultra/parakeet-ultra-F32.gguf \
+  --manifest samples/wer/tedlium-longform.manifest.jsonl \
+  --out reports/wer/parakeet-ultra-F32.tedlium-longform.b1.jsonl
+
+uv run scripts/wer/score.py <report>.jsonl
+```
@@ -4,10 +4,10 @@ NVIDIA's [Parakeet](https://huggingface.co/collections/nvidia/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 an
-80-bin mel frontend; English-only across the family, except
-`parakeet-tdt-0.6b-v3` and its German fine-tune `parakeet-primeline`,
-which cover 25 European languages.
+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` 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
@@ -19,9 +19,11 @@ 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-tdt-0.6b-v3`.** The
-  only multilingual variant; same size as v2, broader coverage at a
-  small English-WER cost.
+- **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](parakeet-ultra.md) for how it segments long audio.
 - **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.
@@ -52,10 +54,11 @@ 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.
 
-<!-- catalog:family variants=parakeet-tdt-0.6b-v2,parakeet-tdt-0.6b-v3,parakeet-primeline,parakeet-tdt-1.1b,parakeet-tdt_ctc-110m,parakeet-tdt_ctc-1.1b,parakeet-rnnt-0.6b,parakeet-rnnt-1.1b,parakeet-ctc-0.6b,parakeet-ctc-1.1b,parakeet-unified-en-0.6b -->
+<!-- catalog:family variants=parakeet-tdt-0.6b-v2,parakeet-ultra,parakeet-tdt-0.6b-v3,parakeet-primeline,parakeet-tdt-1.1b,parakeet-tdt_ctc-110m,parakeet-tdt_ctc-1.1b,parakeet-rnnt-0.6b,parakeet-rnnt-1.1b,parakeet-ctc-0.6b,parakeet-ctc-1.1b,parakeet-unified-en-0.6b -->
 | 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-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-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](parakeet-tdt-0.6b-v3.md) |
 | `parakeet-primeline`       |   627M | 25 languages + auto-detect |    740 MB | FLEURS de (WER)              | 5.98% | token timestamps            | [parakeet-primeline.md](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-1.1b.md) |
@@ -80,7 +83,9 @@ each per-variant doc has direct download links.
 
 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. See
+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](../input-limits.md).
 
 ## Quick start
@@ -118,8 +123,10 @@ 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,
-VAD, speaker diarization. Language coverage is English-only except
-`parakeet-tdt-0.6b-v3` and `parakeet-primeline` (25 European languages,
+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