File size: 4,329 Bytes
a6de3c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a152b32
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a6de3c1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
---
license: apache-2.0
language:
- en
- hi
- ta
- te
- ml
- bn
- mr
- kn
base_model:
- Aratako/MioCodec-25Hz-24kHz
pipeline_tag: audio-to-audio
---
**Overview**
- **What:** Lightweight toolkit to run MioCodec decoding in two modes: streaming (real‑time token→audio from an LLM websocket) and offline (JSONL of embeddings → WAV).
- **Why:** You can avoid pushing the entire MioCodec source to your HF model repo — push only checkpoints and a few small helper files instead.

**Requirements**
- See [MioCodec-stream/requirements.txt](MioCodec-stream/requirements.txt#L1) for the base list. Minimal packages:
  - `torch`, `numpy`
  - `git+https://github.com/Aratako/MioCodec.git` (or `miocodec` package if available)
  - `websockets` (streaming)
  - `soundfile` (offline WAV write) or `sounddevice` (optional playback)

**Repository contents (what to push to HF)**
- `weights/` — your `lora_windowed` checkpoint files (use Git LFS).
- `streaming_istft_head.py` — streaming-capable ISTFT head (small helper).
- `windowed_streamer.py` — windowed streamer implementation.
- `mio_codec_wrapper.py` — wrapper script for streaming and offline decode. See [mio_codec_wrapper.py](mydrive2/Indic-Mio/MioCodec/mio_codec_wrapper.py#L1).
- `requirements.txt` — environment dependencies. See [MioCodec-stream/requirements.txt](MioCodec-stream/requirements.txt#L1).

**Install (inference host)**
1. Create a fresh venv and activate it:

```bash
python3 -m venv ~/mio_env
source ~/mio_env/bin/activate
```

2. Install dependencies:

```bash
pip install -r MioCodec-stream/requirements.txt
```

If you prefer to install MioCodec from the GitHub repo directly:

```bash
pip install git+https://github.com/Aratako/MioCodec.git
```

**Usage — Streaming (LLM websocket → audio)**
- Start/point your LLM (vLLM or other) to stream embeddings to a websocket endpoint (JSON messages). Expected message format (per message):

```json
{"embedding": [0.12, 0.34, ...]}
```

Send a final message `{ "finished": true }` when the utterance completes.

Run the wrapper in streaming mode (plays audio locally if `--play`):

```bash
python mydrive2/Indic-Mio/MioCodec/mio_codec_wrapper.py --mode stream \
  --uri ws://localhost:8000/stream --device cpu --play
```

Notes:
- The wrapper prefers local helper files (`windowed_streamer.py`, `streaming_istft_head.py`) when present; otherwise it will attempt to use the installed `miocodec` package and swap in the streaming head if available.
- Adjust `--chunk-size`, `--left`, `--right` options to tune streaming latency/quality.

**Usage — Offline (JSONL tokens → WAV)**
- Prepare a JSONL file where each line is a JSON object with an `embedding` array, e.g.:

```jsonl
{"embedding": [ ... ]}
{"embedding": [ ... ]}
```

Run offline decode:

```bash
python MioCodec-stream/mio_codec_wrapper.py --mode offline --input tokens.jsonl --out out.wav --device cpu
```

The wrapper will prefer the `windowed_streamer` workflow (exact frame accounting). If the streamer is not available it falls back to a best-effort one-shot decode (less exact).

**Pushing weights to Hugging Face**
- Only push the small helper files and the `weights/` folder to your HF repo. Use Git LFS for large checkpoint files:

```bash
git lfs install
git lfs track "weights/*"
git add .gitattributes weights/ streaming_istft_head.py windowed_streamer.py mio_codec_wrapper.py requirements.txt
git commit -m "Add weights + streaming helpers"
git push
```

When calling `MioCodecModel.from_pretrained()` in the wrapper, you can pass the HF repo id (for example `PRIYANSHUDHAKED/Mio-Codec-Stream`) so the model code downloads the weights from the hub automatically.

**Troubleshooting**
- Import errors (e.g. missing `BloomPreTrainedModel`) usually come from incompatible `transformers` versions — use the version recommended by the model package or install inside a fresh venv.
- If audio shows seams, increase `--right` lookahead or ensure `windowed_streamer.py` is present so exact frame trimming is used.
- If playback is silent but files exist, inspect `out.wav` in a player; the wrapper may write raw float32 if `soundfile` is unavailable.

**Next steps / options**
- I can add a tiny `README` section that shows how to configure a vLLM or other LLM server to emit the expected websocket JSON format. Want that added?