Parakeet TDT 0.6B v3 — LiteRT (TFLite) for Android, on-device
NVIDIA Parakeet TDT 0.6B v3 (25 European languages, punctuation and capitalization) converted to LiteRT / TFLite for fully offline transcription on Android phones, CPU only (XNNPACK). Used by the Recallina voice-notes app.
Converted from the ONNX export istupakov/parakeet-tdt-0.6b-v3-onnx.
Files (≈ 587 MB)
| file | what |
|---|---|
enc_p1_drq.tflite | encoder: subsampling + layers 0–7 · in [1,128,1000] log-mel → out [1,125,1024] |
enc_p2_drq.tflite | encoder layers 8–15 · [1,125,1024] → [1,125,1024] |
enc_p3_drq.tflite | encoder layers 16–23 · [1,125,1024] → [1,1024,125] |
pred_drq.tflite | prediction network (embedding + 2×LSTM 640 + joint projection) · signature tok[1] int32, h[2,640], c[2,640] → pred[1,640], h, c |
joint_drq.tflite | joint network · f[1,1024], pred[1,640] → logits[1,8198] (8193 tokens + 5 TDT durations) |
vocab.txt | SentencePiece vocabulary (8193 tokens, last = blank) |
mel_fb.bin | mel filterbank [257][128] float32 little-endian |
manifest.json | sizes and SHA-256 |
All models are dynamic-range quantized (int8 weights, float activations). The encoder is split in three parts so that only ~200 MB needs to be loaded at a time.
How to run
- Audio: 16 kHz mono. Features as NeMo: pre-emphasis 0.97, reflect pad 256, symmetric Hann window of
400 samples centered in a 512-point FFT, hop 160, power spectrum ×
mel_fb,log(x + 2^-24), per-band normalization over the whole recording (mean, std with n−1, + 1e-5). - Encoder on fixed blocks of 1000 frames (10 s), overlapping by 200 frames: keep encoder frames
[12, 112)of each block ([0, …)for the first, up to the valid length for the last). - Greedy TDT decoding: start with
tok = blank (8192)and zero states; for each kept encoder frame take argmax over tokens and over the 5 durations, emit non-blank tokens (update states), advance by the predicted duration (at most 10 tokens per frame).▁= space.
Quality and speed
13 real Italian voice notes (5 min), word difference vs Whisper large-v3 transcripts: FP32 ONNX 16.9 % · this model 16.1 % · full int8 (weights + activations) 20.9 %. Samsung Galaxy S21 (Exynos 2100), 4 CPU threads at background priority: ≈ 1/3 of real time.
License
CC-BY-4.0, like the original model. Contains NVIDIA Parakeet TDT 0.6B v3 by NVIDIA, modified (converted to LiteRT, quantized, split). Attribution required in apps using it.