---
license: mit
library_name: mnn
tags:
  - text-to-speech
  - indic
  - mnn
  - vocos
  - f5-tts
  - indian-languages
  - quantization
pipeline_tag: text-to-speech
base_model: ai4bharat/IndicF5
---

# IndicF5 MNN

MNN exports of [ai4bharat/IndicF5](https://huggingface.co/ai4bharat/IndicF5) (DiT + Vocos), including fp16 and int8 weight-quantized variants.

## Files

### DiT

| File | Approx. size |
|------|----------------|
| `indicf5_dit_fp32.mnn` | ~1.3 GB |
| `indicf5_dit_fp16.mnn` | ~646 MB |
| `indicf5_dit_int8wq.mnn` | ~347 MB |
| `indicf5_dit_fp16_int8wq.mnn` | ~646 MB |

### Vocos

| File | Notes |
|------|--------|
| `indicf5_vocos_fp32.mnn` | static |
| `indicf5_vocos_fp32_dynamic.mnn` | variable mel frames (recommended for audio) |
| `indicf5_vocos_fp16.mnn` | |
| `indicf5_vocos_int8wq.mnn` | static |
| `indicf5_vocos_int8wq_dynamic.mnn` | variable mel frames |
| `indicf5_vocos_fp16_int8wq.mnn` | |

Sample WAVs (fp32 / int8wq Vocos decode) are under `samples/`. Bench numbers: `quant_benchmark.json`.

## Notes

- Prefer `*_dynamic.mnn` Vocos models for full-length audio (call `resizeTensor` with a **tuple** shape, then `resizeSession`).
- Path used for samples: PyTorch mel → MNN Vocos.
- Still not realtime on Apple M1 CPU for full DiT ODE + Vocos.

## Attribution

Derived from [ai4bharat/IndicF5](https://huggingface.co/ai4bharat/IndicF5). Follow the base model license/terms for redistribution and use.