yogenghodke/indic-parler-tts-mlx-q8

indic-parler-tts — quantized MLX (q8 + bf16)

A fast, 8-bit-quantized MLX build of ai4bharat/indic-parler-tts for Apple Silicon — ~3–4× faster than the unquantized MLX baseline, with no audible quality loss. Speaks English + ~14 Indic languages locally.

Credits 🙏

• Model: AI4Bharat — ai4bharat/indic-parler-tts (Apache-2.0).
• Pure-MLX implementation: Akash Singh (@akashicMarga) — mlx-audio-train. Used with permission.
• Quantization (q8 + bf16): Yogen Ghodke (@yogen-ghodke-113) — contributed upstream in PR #9.

Speed (M4 Mac mini, same sentence)

Config	Realtime factor
CPU PyTorch	0.31×
MLX fp32	0.49×
MLX q8	0.97×
MLX q8 + bf16	1.28×

8-bit beats 4-bit here on both speed and quality at these matrix sizes.

Usage

git clone https://github.com/akashicMarga/mlx-audio-train  # + PR #9

from models.indic_parler_tts import load_model, generate
import soundfile as sf
model, toks = load_model("ai4bharat/indic-parler-tts", quantize=8, bf16=True)
audio = generate(model, toks,
                 description="Sunita speaks in a clear, natural voice at a moderate pace.",
                 text="शिक्षण हे समाज परिवर्तनाचे सर्वात प्रभावी साधन आहे.")
sf.write("out.wav", audio, 44100)

How it works

Quantizes only the AR decoder's nn.Linear layers (the ~89% bottleneck), skipping embeddings (the implementation slices embed_positions.weight directly, so packing it breaks the forward pass); the T5 encoder + DAC stay fp32. bf16 halves activation + KV-cache traffic in the AR loop. Open-sourced for the greater good. ❤️