Does speech-swift work on iOS?

Kokoro TTS, Qwen3.5-Chat, Silero VAD, Parakeet ASR, Omnilingual ASR (300M), DeepFilterNet3, and WeSpeaker all run on iOS 17+ via CoreML on the Neural Engine. MLX-based models (Qwen3-ASR, Qwen3.5-Chat MLX, PersonaPlex, Omnilingual 1B/3B/7B) require macOS 14+ on Apple Silicon. Qwen3-TTS has both MLX and CoreML backends.

Does speech-swift require an internet connection?

Only for the initial model download from HuggingFace. After that, all inference runs fully offline with no network access. No cloud APIs, no API keys needed.

How does speech-swift compare to Whisper?

Qwen3-ASR-0.6B achieves RTF 0.06 on M2 Max — 40% faster than Whisper-large-v3 via whisper.cpp (RTF 0.10) — with comparable accuracy across 52 languages. speech-swift provides a native Swift async/await API, while whisper.cpp requires a C++ bridge.

What Apple Silicon chips are supported?

All M-series chips: M1, M2, M3, M4 and their Pro/Max/Ultra variants. Requires macOS 14+ (Sonoma) or iOS 17+.

Can I use speech-swift in a commercial app?

Yes. speech-swift is licensed under Apache 2.0. The underlying model weights have their own licenses — check each model's HuggingFace page for details.

How much memory does speech-swift need?

From ~3 MB (Silero VAD) to ~6.5 GB (PersonaPlex 7B). Kokoro TTS uses ~200 MB, Qwen3-ASR ~2.2 GB, Qwen3-TTS ~2 GB. Multiple models can run simultaneously using CoreML on the Neural Engine alongside MLX on the GPU.

On-Device Speech SDK

Getting Started — Apple Getting Started — Android

Apple — Homebrew:

brew tap soniqo/speech https://github.com/soniqo/speech-swift && brew install speech

Android — Gradle:

implementation("audio.soniqo:speech:0.0.5")

Speech-to-Text

Text-to-Speech

Language Models

Qwen3.5-0.8B Chat

On-device LLM, hybrid DeltaNet + GatedAttention, INT4 MLX / INT8 CoreML, streaming tokens

MLXCoreML

Speech-to-Speech

PersonaPlex 7B

Full-duplex spoken dialogue, 18 voice presets, RTF ~0.87

MLX

Audio Analysis

Android & Linux

Performance

0.06x

ASR 17x faster than real-time

0.55x

TTS 37ms/step, faster than real-time

0.87x

Speech-to-Speech 68ms/step, full-duplex

~15

LLM tok/s Qwen3.5-0.8B INT4, M-series

23x

Streaming VAD 32ms chunks, sub-ms latency

2.35%

ASR WER Beats Whisper Large v3 Turbo

99.1%

VAD F1 FireRedVAD on FLEURS

View all benchmarks →

Memory

2.2 GB

Qwen3-ASR 0.6B 4-bit MLX · 675 MB weights

2 GB

Qwen3-TTS 0.6B 4-bit MLX · 977 MB weights

500 MB

Kokoro-82M CoreML · 325 MB weights

700 MB

Qwen3.5-Chat 0.8B INT4 MLX · 418 MB weights

1.5 GB

CosyVoice3 0.5B 4-bit MLX · 732 MB weights

6.5 GB

PersonaPlex 7B 4-bit MLX · 4.9 GB weights

400 MB

Parakeet TDT 0.6B CoreML INT8 · 500 MB weights

On-Device Speech SDK

Qwen3-ASR

Parakeet TDT

Omnilingual ASR

Forced Alignment

Qwen3-TTS

CosyVoice3

Kokoro-82M

Voice Cloning

Qwen3.5-0.8B Chat

PersonaPlex 7B

Voice Activity Detection

Wake-Word / KWS

Speaker Diarization

Speaker Embeddings

Speech Enhancement

Source Separation

Parakeet TDT v3

Kokoro-82M

Silero VAD v5

DeepFilterNet3