Modelpublic

sonic-speech/parakeet-tdt-0.6b-v3

sourceHugging Facecc-by-4.0updated 7mo agoView on Hugging Face

1likes52downloads

Model Card

Parakeet TDT 0.6B v3 (MLX, BF16)

NVIDIA Parakeet TDT v3 automatic speech recognition model in MLX BF16 SafeTensors format for Apple Silicon. This is the reference BF16 checkpoint — see quantized variants for reduced memory and faster inference:

—`sonic-speech/parakeet-tdt-0.6b-v3-int8` — Encoder INT8 (recommended)
—`sonic-speech/parakeet-tdt-0.6b-v3-int4` — Encoder INT4 (lite, for 8GB Macs)

Benchmark Results (M3 Max, 64GB)

Variant Comparison

Variant	Size	WER (LibriSpeech)	WER (TED-LIUM)	RTFx	Peak Memory
BF16	1,254 MB	0.82%	15.1%	73x	3,002 MB
INT8	755 MB	0.82%	15.1%	95x	1,268 MB
INT4	489 MB	0.82%	15.5%	98x	1,003 MB

—LibriSpeech test-clean: 50 samples, studio-quality read speech
—TED-LIUM: 8 TED talks (60s segments), real-world acoustics

RTFx vs Audio Duration

5s	15s	30s	60s	120s
59x	91x	98x	103x	99x

Quantization Strategy

Encoder-only mixed-precision: the Conformer encoder (~85% of parameters) is quantized while the decoder and joint network remain BF16. This preserves decoder precision for rare words and punctuation.

Variant	Size Reduction	Speed Gain	Memory Reduction	WER Impact
INT8	-40%	+30%	-58%	None
INT4	-61%	+34%	-67%	+0.4pp on real speech

Usage

Install:

Model Details

—Architecture: Conformer encoder + TDT (Token-and-Duration Transducer) decoder
—Parameters: 627M
—Languages: 25 (SentencePiece tokenizer)
—Sample rate: 16 kHz
—Precision: BF16 (optimized for Apple Silicon)

Origin

Weights from `mlx-community/parakeet-tdt-0.6b-v3`, converted from NVIDIA's official `nvidia/parakeet-tdt-0.6b-v3`.

Part of the Sonic Speech model collection for the Sonic local-first voice AI project.