---
family: parakeet
license: cc-by-4.0
model_id: parakeet-tdt_ctc-0.6b-ja__parakeet_tdt_ctc_0.6b_ja_5s_f32_stateful
source_url: https://huggingface.co/litert-community/parakeet-tdt_ctc-0.6b-ja
task: automatic-speech-recognition
---

# parakeet-tdt_ctc-0.6b-ja__parakeet_tdt_ctc_0.6b_ja_5s_f32_stateful

| | |
|---|---|
| **Task** | automatic-speech-recognition |
| **Family** | parakeet |
| **Source** | https://huggingface.co/litert-community/parakeet-tdt_ctc-0.6b-ja |
| **License** | cc-by-4.0 |

## Conversion

| | |
|---|---|
| **Tool** | litert-torch via the official litert-samples speech_recognition convert pipeline (ParakeetTDT path; ParakeetTDTCTCJa subclass, litert-samples#277) 0.10.0 (editable dev checkout 115a136) |
| **Command** | `python stageA_ja.py && python stageB_ja.py --quant=none --stateful_after=4 --output=parakeet_tdt_ctc_0.6b_ja_5s_f32_stateful.tflite  # two-process split of convert_to_tflite.py --model=nvidia/parakeet-tdt_ctc-0.6b-ja --quant=none --input_sec=5 --stateful_after=4` |
| **Quantization** | none (float32 weights) |

## Artifacts

| File | SHA-256 | Size (MB) |
|---|---|---|
| `parakeet_tdt_ctc_0.6b_ja_5s_f32_stateful.tflite` | `b787283b9f4b823e14f312143d1c0e4ae0b02d63efc2c6baf793065412347207` | 2280.366 |

## Performance

No benchmark data yet.

## Device runs (NPU / LiteRT-LM)

Model-level on-device results — measured behavior of this exact artifact on the recorded device, only meaningful together with that environment. Throughput is conditional on the measured prompt length (Prompt (tokens); '-' = the source stated none): rows differing only there are different measurements, not re-runs. Full records (failure classes, log evidence): `data/device_runs/2.2.0/2026-08-27/parakeet-tdt_ctc-0.6b-ja__parakeet_tdt_ctc_0.6b_ja_5s_f32_stateful__galaxy-s26.json`.

| Device | Accelerator | Status | Full delegation | Output match | Prompt (tokens) | Latency p50 (ms) | Prefill tok/s | Decode tok/s | TTFT (ms) | Peak mem (MB) | Environment | Date | Provenance |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| galaxy-s26 | gpu_mldrift | pass | - | - | - | 63.033 | - | - | - | - | Galaxy S26 (SM-S942Q) · Qualcomm SM8850 · litert 2.2.0 · Android 16 | 2026-08-27 | measured |
| galaxy-s26 | npu_qnn | pass | - | - | - | 30.315 | - | - | - | - | Galaxy S26 (SM-S942Q) · Qualcomm SM8850 · litert 2.2.0 · QAIRT (Hexagon, AOT host-compiled, SM8850 target) · Android 16 | 2026-08-27 | measured |

## Pitfalls

- Three signatures: encode (log-mel [1,80,500] -> [1,1024,63]) + decode (4-token stateless prefix; N=4 matches the official parakeet-tdt-0.6b-v3 stateful layout) + decode_1 (single-step, LSTM h/c carried between calls; logits [1,63,64,3078] = 3072 tokens + blank 3072 + 5 durations). 80 mel bins (v3 uses 128); NeMo preprocessing: preemph 0.97, n_fft 512, win 25 ms, hop 10 ms, per-feature norm (measured; HF card).
- Exact conversion: 28/28 five-second windows reproduce NeMo transcribe() (joined CER 0.0) on a 137 s CC0 test read; Pixel 8a GPU fully delegates both signatures (1862/1862 + 2083/2083 LITERT_CL) with identical greedy ids — encode 239 ms, decode 95 ms/call, ~20 calls per 5 s window (measured 2026-08-13).
- 2.4 GB float32 file — GPU compile on the Pixel 8a takes 25-37 s; the i8 sibling is 608 MB but fails Mali GPU compile (see the i8 meta) (measured).
- NeMo trap for anyone re-deriving references: model.transcribe() leaves decoder/joint in training mode (dropout on) — compute reference tensors before transcribe(), or re-.eval() after (measured; caused a false parity failure).
- Mali fp16 quirk: decode logits at token positions >= ~43 go NaN when the token suffix is zero-padded garbage (probe fixture artifact); real-prefix cells are clean and greedy decoding is unaffected (measured).
- Stateful speed (measured 2026-08-13): Pixel 8a GPU compile 7.6 s, decode 16 ms, decode_1 8 ms per call — a 5 s window drops from ~2.1 s (stateless file) to ~0.45 s; the 28-window desktop sweep runs 6x faster (18.2 s -> 3.0 s).
- The litert-samples TdtDecoder.kt stateful path has two decoding bugs (state adoption on blank steps; decode->decode_1 switch one token early) that cost accuracy (16/28 windows vs NeMo). With corrected RNN-T semantics this stateful lineage reproduces NeMo 28/28 f32 (CER 0.0000). Fix proposed as litert-samples#278; until merged prefer the stateless files with the stock app (measured; HF card 'Stateful decoding' section).

---

_Rendered by `edge-card` from `card.json` (the source of truth); edit the inputs, not this file._