{ "artifacts": [ { "file": "parakeet_tdt_ctc_0.6b_ja_5s_f32.tflite", "sha256": "db103d352034fba5dfb43a8cbba9dba72afecc730ec5f9999200c6ea41a27a9b", "size_mb": 2280.836 } ], "benchmarks": [], "browser": { "backends": [ { "backend": "wasm_xnnpack", "date": "2026-08-13", "env": { "browser": "chromium", "browser_version": "151.0.7922.34", "headless": true, "jspi": true, "litertjs_core_version": "2.5.3", "machine_label": "mac-studio-m4-max", "os": "macOS", "os_version": "27.0.0", "webgpu_adapter": { "architecture": "metal-3", "description": "", "device": "", "vendor": "apple" } }, "full_delegation": null, "latency_p50_ms": null, "loads": false, "max_rel_diff": null, "output_match": null, "provenance": "measured", "runs": false }, { "backend": "webgpu_mldrift", "date": "2026-08-13", "env": { "browser": "chromium", "browser_version": "151.0.7922.34", "headless": true, "jspi": true, "litertjs_core_version": "2.5.3", "machine_label": "mac-studio-m4-max", "os": "macOS", "os_version": "27.0.0", "webgpu_adapter": { "architecture": "metal-3", "description": "", "device": "", "vendor": "apple" } }, "full_delegation": null, "latency_p50_ms": null, "loads": false, "max_rel_diff": null, "output_match": null, "provenance": "measured", "runs": false } ], "demo_url": null, "sweep_source": "data/sweep/2.5.3/2026-08-13/parakeet-tdt_ctc-0.6b-ja__parakeet_tdt_ctc_0.6b_ja_5s_f32.json" }, "conversion": { "command": "python stageA_ja.py && python stageB_ja.py --quant=none --output=parakeet_tdt_ctc_0.6b_ja_5s_f32.tflite # two-process split of convert_to_tflite.py --model=nvidia/parakeet-tdt_ctc-0.6b-ja --quant=none --input_sec=5", "quantization": "none (float32 weights)", "tool": "litert-torch via the official litert-samples speech_recognition convert pipeline (ParakeetTDT path; ParakeetTDTCTCJa subclass, litert-samples#277)", "tool_version": "0.10.0 (editable dev checkout 115a136)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "gpu_mldrift", "context_length": null, "date": "2026-08-27", "decode_tokens_per_s": null, "delegated_ops": null, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-npubench", "os_build": "Android 16", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "npubench sweep row: N=50 median, 1 backend = 1 process, accepted only with thermal NONE->NONE; NPU rows additionally required qnn_partition delegate evidence in logcat (a stock model asked for on the NPU can silently land on XNNPACK and report a CPU number)", "mode=jit thermal=NONE->NONE headroom=0.71333337->0.72333336 attempt=0" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": 63.606, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "iterations": 50, "latency_max_ms": 67.819, "latency_min_ms": 62.363, "load_ms": 5984.2 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-08-27/parakeet-tdt_ctc-0.6b-ja__parakeet_tdt_ctc_0.6b_ja_5s_f32__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "npu_qnn", "context_length": null, "date": "2026-08-27", "decode_tokens_per_s": null, "delegated_ops": null, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-npubench", "os_build": "Android 16", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Qualcomm SM8850", "vendor_sdk": "QAIRT (Hexagon, AOT host-compiled, SM8850 target)" }, "error": null, "evidence": [ "npubench sweep row: N=50 median, 1 backend = 1 process, accepted only with thermal NONE->NONE; NPU rows additionally required qnn_partition delegate evidence in logcat (a stock model asked for on the NPU can silently land on XNNPACK and report a CPU number)", "mode=aot thermal=NONE->NONE headroom=0.70333326->0.71 attempt=0" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": 30.326, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "iterations": 50, "latency_max_ms": 30.786, "latency_min_ms": 29.873, "load_ms": 419.7, "qnn_partition_lines": 61 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-08-27/parakeet-tdt_ctc-0.6b-ja__parakeet_tdt_ctc_0.6b_ja_5s_f32__galaxy-s26.json" } ] }, "model": { "family": "parakeet", "id": "parakeet-tdt_ctc-0.6b-ja__parakeet_tdt_ctc_0.6b_ja_5s_f32", "license": "cc-by-4.0", "source_url": "https://huggingface.co/litert-community/parakeet-tdt_ctc-0.6b-ja", "task": "automatic-speech-recognition" }, "pitfalls": [ "Two signatures: encode (log-mel [1,80,500] -> [1,1024,63]) + decode (stateless 64-token TDT loop; 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)." ], "schema_version": "1.2" }