moonhouse/pianissimo-sv-mlx

Klang Pianissimo — MLX

An MLX conversion of Klang Pianissimo by Klang AI AB, a Swedish speech recognition model (FastConformer encoder + TDT decoder, fine-tuned from NVIDIA Parakeet TDT 0.6B v3), optimized for on-device inference on Apple Silicon (M1/M2/M3/M4) via parakeet-mlx.

All credit for the model goes to Klang AI AB. This repository only provides the weights formatted for Apple Silicon MLX; the weights were not retrained. Licensed under CC BY 4.0, like the original.

---

⚡ Quickstart

1. Installation

pip install parakeet-mlx

(Ensure ffmpeg is installed: brew install ffmpeg)

2. Usage in Python

from parakeet_mlx import from_pretrained

# Load model from Hugging Face directly into Apple Silicon unified memory
model = from_pretrained("moonhouse/pianissimo-sv-mlx")

# Transcribe any Swedish audio file (.wav, .mp3, .m4a, etc.)
result = model.transcribe("audio.wav")
print(result.text)

# Access word-level timestamps
for sentence in result.sentences:
    for token in sentence.tokens:
        print(f"[{token.start:.2f}s -> {token.end:.2f}s] {token.text}")

---

📊 Evaluation & Benchmark

Benchmarked on Apple Silicon GPU vs PyTorch NeMo CPU on the Swedish Speech Benchmark (KlangAI/klang-dialects):

Metric	PyTorch NeMo (CPU)	Apple Silicon MLX (GPU)
Word Error Rate (WER)	4.67%	6.67%
Model-to-Model Agreement	—	98.00%
Inference Speedup	1.0x (baseline)	~8.8x faster
Model Load Time	~7.6s	~0.06s (instant zero-copy)

---

📁 Repository Files

File	Purpose
model.safetensors	Transformed FastConformer & TDT weights (MLX memory layout, float16)
config.json	Architecture specification & local attention configuration
tokenizer.model	SentencePiece BPE tokenizer

---

📜 Attribution & License

• Original Model: KlangAI/pianissimo-sv by Klang AI AB.
• License: CC BY 4.0.