NVIDIA Parakeet TDT 0.6B v2

NVIDIA Parakeet TDT 0.6B v2 is a 600-million-parameter automatic speech recognition (ASR) model designed exclusively for English speech transcription. Released in May 2025, it secured the top spot on the Hugging Face Open ASR leaderboard with an average Word Error Rate (WER) of 6.05%, outperforming comparably sized open ASR models at the time of its release. The model is built by NVIDIA under the CC-BY-4.0 license, making it freely usable for both research and commercial applications.

This is not a general-purpose language model or a multimodal system — it is a purpose-built transducer for converting audio to text. Its dense architecture, combined with the Token-and-Duration Transducer (TDT) decoder, allows it to process an hour of audio in roughly one second under optimal batch configurations. For developers and engineers who need local, offline transcription with high accuracy and low latency, Parakeet TDT 0.6B v2 offers a compelling balance of model size and performance.

Parakeet TDT 0.6B v2 uses a FastConformer encoder paired with a Token-and-Duration Transducer (TDT) decoder. This is a dense model — all 0.6B parameters are active during every inference pass. There are no expert routing or sparse activation mechanisms. That means the VRAM footprint is predictable and scales linearly with precision and batch size.

The model ingests raw audio (typically 16 kHz mono WAV or FLAC) and outputs text with word-level timestamps, punctuation, and capitalization. It does not require a separate language model or text post-processing pipeline. The TDT decoder jointly predicts token sequences and their durations, enabling precise timing information for each transcribed word.

Key architectural specs:

• Parameters: 600M (dense)
• Encoder: FastConformer (efficient streaming-friendly variant of Conformer)
• Decoder: Token-and-Duration Transducer (TDT)
• Input: Audio (16 kHz, mono)
• Output: Text with word-level timestamps and punctuation
• Maximum segment length: Up to 24 minutes of audio in a single forward pass

The model’s small footprint and efficient encoder make it suitable for real-time and near-real-time workloads. On the Hugging Face Open ASR leaderboard, it achieved a Real-Time Factor (RTF) of 3380 with a batch size of 128 — meaning it processes audio 3380× faster than real time under those conditions.

This model is built for one job: turning spoken English into written text, accurately and fast. It does not support other languages, generate speech, or answer questions. But within its domain, it delivers strong results across a variety of acoustic environments.

Benchmarked WER on standard datasets (from HuggingFace model card):

The model handles clean read speech (LibriSpeech) with near-perfect accuracy and degrades gracefully in noisy, conversational, or financial earnings call settings. It also performs well on challenging content like spoken numbers and song lyrics, thanks to the TDT decoder’s ability to model duration patterns.

Concrete use cases:

• Real-time transcription for meetings and live events – low latency (160ms minimum in streaming mode) and word timestamps make it suitable for live captioning.
• Archival audio transcription – process hours of recorded content quickly offline, without cloud dependencies.
• Voice-controlled local applications – integrate into on-device voice assistants or dictation tools that must run entirely on consumer hardware.
• Podcast and media production – generate accurate transcripts with punctuation, ready for subtitle files or search indices.

The model automatically inserts punctuation and capitalizes proper nouns, so the output is readable without a separate formatting step.

Because Parakeet TDT 0.6B v2 is a 600M-parameter dense model, it runs comfortably on consumer-grade GPUs — and even on some CPUs for offline batch processing. The primary framework for inference is NVIDIA NeMo, but the model is also available through HuggingFace transformers and can be loaded with pipeline("automatic-speech-recognition").

Important: ASR models are typically run in FP16 or FP32. INT8 and GGUF quantization support for ASR is less mature than for LLMs. For most local deployments, FP16 on a GPU is the safe and performant choice.

• RTX 3060 (12 GB) or higher – can run FP16 with large batch sizes for offline processing.
• RTX 4090 (24 GB) – overkill for single-stream inference, but ideal for batch processing dozens of audio files simultaneously.
• Apple Silicon (M2 Pro/Max, M4 Max) – runs well via NeMo’s MPS backend or via transformers with FP16. Expect near real-time speed on M4 Max.
• CPU-only (16+ GB RAM) – possible using transformers with FP32, but speed will be roughly 1–2× real-time (i.e., processing a 1-hour file in 30–60 minutes).

• Real-Time Factor (RTF): 0.0003 (batch=128) to 0.02 (batch=1) on a modern GPU (e.g., RTX 4090). This means transcribing a 1-hour audio file takes <2 seconds with optimal batching, or about 1 minute of wall time for streaming single-sentence segments.
• Latency: Streaming mode achieves 160ms minimum end-to-end latency (first word output).
• Throughput: With a batch size of 128, you can process roughly 20 hours of audio per minute on an RTX 4090.

1import torch
2from transformers import pipeline
3
4pipe = pipeline(
5    "automatic-speech-recognition",
6    model="nvidia/parakeet-tdt-0.6b-v2",
7    device="cuda" if torch.cuda.is_available() else "cpu",
8    return_timestamps="word"
9)
10
11result = pipe("audio_file.flav")
12print(result["text"])

For streaming or low-latency use cases, refer to the NVIDIA NeMo documentation for the streaming inference API.

In the open ASR landscape, Parakeet TDT 0.6B v2 competes most directly with OpenAI Whisper small (244M params) and Whisper medium (769M params). All three are English ASR models, but their architectures and trade-offs differ.

*Approximate from leaderboard data at time of Parakeet’s release.

When to choose Parakeet TDT 0.6B v2:

• You need word-level timestamps without post-processing.
• You want punctuation and capitalization out of the box.
• You prioritize speed — it is 2–5× faster than Whisper medium at similar accuracy.
• You are running on consumer hardware with limited VRAM (Whisper medium requires ~3 GB in FP16, while Parakeet uses ~1.2 GB).

When to choose Whisper small or medium:

• You need multilingual transcription (Whisper supports 99 languages).
• You need to run on CPU without optimization (Whisper’s ONNX runtime has wider CPU support).
• You want a more mature ecosystem for fine-tuning and deployment (Whisper has broader community tooling).

For English-only, high-speed, low-VRAM local ASR, Parakeet TDT 0.6B v2 is currently the best option at its size. The 5–10× speed advantage over Whisper medium, combined with built-in formatting, makes it the practical choice for developers who need to transcribe large volumes of English audio on a single GPU.