WavLM Base Plus MLX fp16
MLX-compatible safetensors export of microsoft/wavlm-base-plus. This bundle is used by speech-swift as the SSL feature-extractor companion for Indic-Mio raw-reference voice cloning.
WavLM runs on 16 kHz mono speech audio and produces hidden states that downstream speech models can use for speaker, content, and paralinguistic representations. For Indic-Mio, speech-swift averages hidden layers 1 and 2, then feeds those features into MioCodec's global encoder to produce the 128-dimensional speaker embedding used by the decoder.
Part of soniqo.audio, an on-device speech toolkit.
Model
| Field | Value |
|---|---|
| Base model | microsoft/wavlm-base-plus |
| Architecture | WavLM / HuBERT-style SSL encoder |
| Parameters | 94M class |
| Format | MLX-compatible safetensors |
| Quantization | None |
| Precision | fp16-compatible safetensors export |
| Sample rate | 16 kHz |
| Hidden size | 768 |
| Layers | 12 Transformer layers |
| Attention heads | 12 |
| Runtime use | Indic-Mio raw-reference speaker embedding companion |
Files
| File | Size | Description |
|---|---|---|
model.safetensors | 360 MB | WavLM weights in safetensors format |
config.json | 2.2 KB | WavLM model configuration |
preprocessor_config.json | 215 B | 16 kHz audio feature-extractor metadata |
soniqo_manifest.json | 622 B | Runtime/export metadata |
Performance
| Check | Result |
|---|---|
| Weight load | Passes in speech-swift WavLM companion loader |
| Indic-Mio raw-reference embedding | Produces finite 128-dimensional MioCodec global embedding |
| Indic-Mio raw-reference synthesis | Passes local E2E synthesis test |
| Hindi ASR sanity on generated clone sample | Recovered नमस्ते यह संदर्भा आवाज़ |
Usage
Swift
import IndicMioTTS
let model = try await IndicMioTTSModel.fromPretrained()
let audio = try await model.generate(
text: "नमस्ते, यह संदर्भ आवाज़ का परीक्षण है। <happy>",
language: "hindi",
referenceAudio: referenceSamples,
referenceSampleRate: referenceSampleRate
)
The runtime downloads this companion automatically when raw reference audio is used. For local testing with a pre-downloaded bundle:
export INDIC_MIO_WAVLM_BUNDLE=/path/to/WavLM-Base-Plus-MLX-fp16
CLI
speech speak \
--engine indic-mio \
--voice-sample reference.wav \
--output clone.wav \
"नमस्ते, यह संदर्भ आवाज़ का परीक्षण है। <happy>"
Source
Converted from microsoft/wavlm-base-plus. The upstream WavLM source project is available from microsoft/unilm.
Links
- speech-swift — Apple SDK
- Speech Studio — local speech generation and voice cloning app
- Docs — install and CLI docs
- soniqo.audio — website
- blog — blog
License
MIT, following the upstream WavLM source project license.