Parakeet-TDT-0.6B-v3-LiteRT-INT8

Parakeet TDT 0.6B v3 — LiteRT (INT8)

NVIDIA's multilingual FastConformer ASR. 25 languages, INT8 encoder + FP32 decoder-joint.

Part of the soniqo.audio speech toolkit — an open, runtime-portable stack for speech AI. This bundle is the LiteRT export, designed to plug into the abstract interfaces in speech-core (C++ voice-agent orchestration library). Browse all LiteRT bundles in the soniqo LiteRT collection.

Use cases on soniqo.audio

• Meeting transcription
• Long-form transcription

NVIDIA's multilingual FastConformer ASR model exported to LiteRT for Android. Covers 25 languages. Split into a FastConformer encoder and a streaming LSTM decoder-joint for token-level inference.

Model

Files

Pipeline

audio [1, N] ──► mel fbank (128 bins, 16 kHz) ──► encoder ──► encoded [1, 1024, T']
                                                                  │
                                                                  ▼
targets (blank-initialized) ──► decoder-joint ──► logits [1, 1, 1, 1030]
                                                         │
                                                         ▼
                                                    TDT decode

TDT (Token-and-Duration Transducer) emits both a token and a duration in {0, 1, 2, 3, 4} frames. Blank id = 1024, vocab size = 1024, total logits = 1030 (1024 tokens + 1 blank + 5 durations).

Encoder signature

Inputs:
  audio_signal  [1, 128, T]   float32   Mel features (log, normalized)
  length        [1]           int64     Valid T (NeMo convention)

Outputs:
  encoded       [1, 1024, T'] float32   Encoded features
  encoded_length [1]          int64     Valid T'

Decoder-joint signature

Inputs:
  encoder_out   [1, 1, 1024]  float32   Current encoder frame
  target        [1, 1]        int64     Last emitted token (blank to start)
  h             [2, 1, 640]   float32   LSTM hidden state
  c             [2, 1, 640]   float32   LSTM cell state

Outputs:
  logits        [1, 1, 1, 1030] float32 Joint output
  h_out         [2, 1, 640]   float32   Next hidden state
  c_out         [2, 1, 640]   float32   Next cell state

Audio preprocessing

The model expects the exact NeMo mel pipeline: 128 mel bins, 16 kHz, n_fft=512, hop_length=160, win_length=400, pre_emphasis=0.97, log mel with per-utterance normalization. Implement this on the caller side in native code to match the NeMo reference exactly.

Source

Upstream: nvidia/parakeet-tdt-0.6b-v3 (CC BY 4.0). 25-language multilingual ASR.

Links

• speech-android — Android SDK
• soniqo.audio — website
• blog — blog

Ecosystem

• soniqo.audio — use-case explorer (transcription, voice cloning, live ASR, voice agents).
• speech-core — C++ orchestration library for voice agents. Abstract STTInterface / TTSInterface / VADInterface / EnhancerInterface; LiteRT implementations plug straight into the interfaces.
• speech-swift — Apple Silicon MLX companion runtime (model-specific MLX bundles linked above where applicable).
• speech-android — Android SDK consuming on-device LiteRT bundles.

Other LiteRT models in this collection

ASR / Transcription

• Nemotron Speech Streaming 0.6B — LiteRT
• Omnilingual ASR CTC 300M — LiteRT
• Omnilingual ASR CTC 300M — LiteRT (INT8)
• Qwen3 ASR 0.6B Encoder — LiteRT (INT8)

VAD / Diarization

• Silero VAD v5 — LiteRT
• Pyannote Segmentation 3.0 — LiteRT
• WeSpeaker ResNet34-LM — LiteRT

TTS / Voice Cloning

• VoxCPM2 — LiteRT (INT8)

License

This bundle inherits the upstream model license (cc-by-4.0). See the linked base_model repository for the full terms.