Indic-Mio-MLX-fp16
Apple Silicon fp16 export of SPRINGLab/Indic-Mio, bundled with the MioCodec 25 Hz / 24 kHz companion required to decode Indic-Mio content tokens into waveform audio.
This is an artifact bundle for Swift/MLX runtime work. It preserves upstream key names; runtime code is expected to implement model loading, sampling, MioCodec conditioning, and waveform decoding.
Model
| Field | Value |
|---|---|
| Primary source | SPRINGLab/Indic-Mio |
| Primary source revision | 25feace00ca76c71c40b1e8d921fd3c2943c545e |
| Bundled codec source | Aratako/MioCodec-25Hz-24kHz |
| Bundled codec revision | 3a737f0de2c6324cb2fe40c1fbd1056c7add423d |
| Format | MLX fp16 safetensors |
| Primary license | Apache-2.0 |
| Codec license | MIT |
| Total bundle size | 1.5 GB |
| Sample rate | 24,000 Hz |
| Voice conditioning | MioCodec global embeddings / zero-shot voice path |
| Runtime status | Artifact export; Swift/MLX runtime port required |
Emotion Control
| Field | Value |
|---|---|
| Marker syntax | Suffix angle tags |
| Supported markers | <happy>, <sad>, <angry>, <disgust>, <fear>, <surprise> |
Files
| File | Size | Description |
|---|---|---|
README.md | 2.3 KB | Model card. |
added_tokens.json | 309.6 KB | Tokenizer asset copied from the upstream Indic-Mio repository. |
bundle_config.json | 697 B | Soniqo bundle metadata linking the text model and bundled codec. |
chat_template.jinja | 205 B | Upstream generation/template metadata. |
config.json | 1.4 KB | Upstream Qwen/Indic-Mio model config at repo root for runtime loading and HF download tracking. |
generation_config.json | 183 B | Upstream generation/template metadata. |
merges.txt | 1.7 MB | Tokenizer asset copied from the upstream Indic-Mio repository. |
miocodec/config.json | 230 B | MioCodec runtime/export config. |
miocodec/config.yaml | 2.4 KB | MioCodec runtime/export config. |
miocodec/model.safetensors | 261.6 MB | Bundled MioCodec fp16 weights required for waveform decoding. |
miocodec/soniqo_manifest.json | 1.3 KB | Export manifest with source revisions, license posture, and bundle metadata. |
model.safetensors | 1.2 GB | Indic-Mio fp16 weights with upstream key names preserved. |
soniqo_manifest.json | 3.1 KB | Export manifest with source revisions, license posture, and bundle metadata. |
special_tokens_map.json | 613 B | Tokenizer asset copied from the upstream Indic-Mio repository. |
tokenizer.json | 13.8 MB | Tokenizer asset copied from the upstream Indic-Mio repository. |
tokenizer_config.json | 2.3 MB | Tokenizer asset copied from the upstream Indic-Mio repository. |
vocab.json | 2.8 MB | Tokenizer asset copied from the upstream Indic-Mio repository. |
Validation
| Test | Result |
|---|---|
| Indic-Mio + MioCodec local roundtrip | Passed |
| Input text | नमस्ते, मैं आज बहुत खुश हूँ। <happy> |
| ASR transcript | नमस्ते, मैं आज बहुत खुश हूँ |
| Sample rate | 24000 Hz |
The roundtrip test was a local synthesis and transcription sanity check. It verifies that the exported Indic-Mio weights and bundled MioCodec path can produce intelligible Hindi speech for an explicit emotion marker. It is not a full runtime benchmark.
Usage
These artifacts are not a drop-in Transformers checkpoint. They are intended for speech-swift or another MLX runtime that implements the Indic-Mio token generation path and MioCodec decoding path.
git lfs install
git clone https://huggingface.co/aufklarer/Indic-Mio-MLX-fp16
Expected bundle layout:
Indic-Mio-MLX-fp16/
model.safetensors
config.json
tokenizer.json
bundle_config.json
soniqo_manifest.json
miocodec/
model.safetensors
config.json
config.yaml
soniqo_manifest.json
Source
- Indic-Mio upstream model: https://huggingface.co/SPRINGLab/Indic-Mio
- MioCodec upstream model: https://huggingface.co/Aratako/MioCodec-25Hz-24kHz
Links
- speech-swift - Apple SDK
- Speech Studio - local speech generation and voice cloning app
- Docs - install and CLI docs
- soniqo.audio - website
- Blog - blog