FluidInference/parakeet-unified-en-0.6b-coreml

🧃 Parakeet Unified EN 0.6B - CoreML

A CoreML-optimized Parakeet Unified 0.6B English ASR model for Apple platforms. A single FastConformer-RNNT checkpoint serves both offline batch transcription and low-latency streaming, with punctuation and capitalization. The encoder is multi-context trained, so each streaming latency tier is the same weights re-exported with a different chunked-attention mask.

This model is consumed by the FluidAudio Swift SDK.

Model Description

• Architecture: FastConformer encoder + RNNT (Transducer) decoder/joint
• Parameters: ~0.6B
• Language: English
• Sample rate: 16 kHz mono
• Vocabulary: 1024 SentencePiece tokens (blank index 1024)
• Subsampling factor: 8 (one encoder frame ≈ 80 ms)
• Precision: FLOAT16; an INT8 encoder variant is included for higher throughput
• Punctuation & capitalization: yes

The same checkpoint exposes two paths:

• Offline batch — full-attention 15 s windows (240,000 samples), long files transcribed with overlapping windows merged on a 2 s overlap.
• Streaming — a [left, chunk, right] chunked-attention window baked into the encoder at conversion time. Latency = (chunk + right) × 80 ms.

Streaming latency tiers

Variant	[L,C,R]	Latency	Aggregate WER*	RTFx	Notes
parakeet-unified-320ms	70, 2, 2	0.32 s	2.37%	10x	lowest latency
parakeet-unified-640ms	70, 7, 1	0.64 s	2.40%	27x	efficiency — same WER as 320 ms, big chunk re-encodes less often
parakeet-unified-1120ms	70, 7, 7	1.12 s	2.25%	33x	best streaming WER
parakeet-unified-2080ms (default)	70, 13, 13	2.08 s	2.47%	54x	default

*Aggregate WER on a 150-file LibriSpeech test-clean sweep (identical files across tiers, so the relative ordering is what matters). Rules of thumb: look-ahead (right context) drives WER, chunk size drives RTFx, and look-ahead saturates around right ≈ 2 frames.

Benchmarks

LibriSpeech test-clean, all 2620 files, INT8 encoder, scored with FluidAudio's TextNormalizer:

Mode	Avg WER	Aggregate WER	Median RTFx	Overall RTFx
batch	2.16%	1.68%	130.0x	143.6x
streaming (2080 ms)	2.21%	1.79%	66.2x	65.9x

vs Parakeet TDT v3 (same harness)

Model	Mode	Avg WER	Overall RTFx	Punctuation/caps	Languages
Parakeet TDT v3	batch (sliding window)	2.6%	110	no	25 + Japanese
Parakeet Unified	batch	2.16%	144	yes	English
Parakeet Unified	streaming	2.21%	66	yes	English

For English transcription, Unified batch beats TDT v3 on both WER and throughput and adds punctuation/capitalization. TDT v3 remains the choice for non-English audio.

Usage in Swift

import FluidAudio

// Offline batch transcription
let manager = try await UnifiedAsrManager.fromHub()
let result = try await manager.transcribe(samples)   // 16 kHz mono Float
print(result.text)

// Streaming (choose a latency tier)
let streaming = try await StreamingUnifiedAsrManager.fromHub(variant: .parakeetUnified2080ms)

The CLI selects a tier with --parakeet-variant, e.g. parakeet-unified-320ms, parakeet-unified-1120ms, parakeet-unified-2080ms (default), or parakeet-unified-offline-15s. See the FluidAudio repository for full instructions.

Repository contents

CoreML pipeline (.mlmodelc), driven from metadata.json / vocab.json:

Stage	Bundle
Preprocessor	parakeet_unified_preprocessor.mlmodelc
Mel encoder (offline)	parakeet_unified_mel_encoder.mlmodelc
Mel encoder (streaming)	parakeet_unified_mel_encoder_streaming_70_13_13.mlmodelc
Encoder (offline)	parakeet_unified_encoder.mlmodelc, _int8
Encoder (streaming)	parakeet_unified_encoder_streaming_{70_2_2,70_7_1,70_7_7,70_13_13}.mlmodelc (+ _int8)
Decoder	parakeet_unified_decoder.mlmodelc
Joint	parakeet_unified_joint.mlmodelc
Joint decision (single step)	parakeet_unified_joint_decision_single_step.mlmodelc

metadata.json holds the runtime shapes (mel/encoder/decoder/state) and streaming window geometry; vocab.json is the 1024-token SentencePiece vocabulary.

License

Released under the CC-BY-4.0 license.

Acknowledgments

Based on NVIDIA's Parakeet (FastConformer-RNNT) model. CoreML conversion and Swift integration by the FluidInference team.