3b6fd99989
WAI is a container + capability-dispatch standard for media. It does not re-implement codecs; it dispatches to SOTA standard libraries (PNG, FLAC, zstd as the mandatory floor; AVIF, JPEG-XL, Opus, AV1, XZ as the recommended modern set) and adds an envelope that lets a neural shared-prior path coexist with the model-free floor. The capability a file requires is named, not supplied. Two paths open every conforming file: the neural condition (sink regenerates from a shared ambient prior) and the zeroth condition (registered SOTA codec menu). Apache-2.0. Adds: - wai/SPEC.md — WAI v1.0 spec (envelope, manifest, capability dispatch, registered capabilities including v1.1 multi-rendition extension) - wai/wai-rs/ — Rust reference implementation: lib + cdylib + staticlib, C FFI, 11 capability + envelope round-trip tests passing - wai/wai-web/ — wasm-bindgen wrapper for browser sinks - wai/wai-neural/ — Python tooling for neural-condition models - wai/tools/ — codec export + encode helpers (Python) - wai/corpus/CORPUS.md — third-party test corpus reacquisition guide Updates README.md to list four standards instead of three. Known consumer: CommunicationOS uses WAI for media transport over MoQT/MLS sessions. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
94 lines
3.1 KiB
Python
94 lines
3.1 KiB
Python
"""WAI source-side neural encoder for `wai.neural.mimi`.
|
||
|
||
Mimi (Kyutai) at 24 kHz, 12.5 Hz frame rate, 8 codebooks of 2048.
|
||
Designed for real-time speech LLM use cases. Token rate is among the
|
||
lowest of any high-quality neural audio codec (~1.1 kbps), which makes
|
||
it the right pick when the wire-byte budget dominates.
|
||
|
||
Wire format (uniform with all wai.neural.<audio> capabilities):
|
||
<III> q t n_samples (little-endian u32 × 3)
|
||
q*t × u16 codes (row-major, LE)
|
||
|
||
Usage:
|
||
python tools/wai_mimi_encode.py <input.wav> -o <output.wai>
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import argparse
|
||
import json
|
||
import struct
|
||
import subprocess
|
||
import sys
|
||
import wave
|
||
from pathlib import Path
|
||
|
||
import numpy as np
|
||
|
||
MODEL_ID = "kyutai/mimi"
|
||
WAI_MAGIC = b"WAI1"
|
||
SR = 24_000
|
||
|
||
|
||
def encode(wav_path: str, out_path: str) -> None:
|
||
with wave.open(wav_path, "rb") as w:
|
||
sr_in = w.getframerate()
|
||
raw_bytes = w.getnframes() * 2 * w.getnchannels()
|
||
|
||
pcm = subprocess.run(
|
||
["ffmpeg", "-v", "error", "-i", wav_path, "-ac", "1",
|
||
"-ar", str(SR), "-f", "f32le", "-"],
|
||
check=True, capture_output=True).stdout
|
||
raw = np.frombuffer(pcm, np.float32)
|
||
|
||
import torch
|
||
from transformers import MimiModel, AutoFeatureExtractor
|
||
model = MimiModel.from_pretrained(MODEL_ID).eval()
|
||
proc = AutoFeatureExtractor.from_pretrained(MODEL_ID)
|
||
|
||
inp = proc(raw_audio=raw, sampling_rate=SR, return_tensors="pt")
|
||
with torch.no_grad():
|
||
enc = model.encode(inp["input_values"], inp.get("padding_mask"))
|
||
# MimiEncoderOutput.audio_codes: [batch=1, n_codebooks, T]
|
||
codes = enc.audio_codes.squeeze(0).numpy().astype(np.uint16)
|
||
q, t = codes.shape
|
||
|
||
payload = struct.pack("<III", q, t, len(raw)) + codes.tobytes()
|
||
|
||
manifest = {
|
||
"wai": "1.0",
|
||
"media": "audio",
|
||
"intent": "replicate",
|
||
"model_requirement": {
|
||
"capability": "wai.neural.mimi",
|
||
"fallback": "wai.audio.opus",
|
||
},
|
||
"conditioning": {"kind": "mimi_tokens"},
|
||
"target": {"sr": SR, "n_samples": int(len(raw)),
|
||
"codebooks": int(q), "frames": int(t)},
|
||
}
|
||
mb = json.dumps(manifest, separators=(",", ":")).encode()
|
||
with open(out_path, "wb") as f:
|
||
f.write(WAI_MAGIC)
|
||
f.write(struct.pack("<I", len(mb))); f.write(mb)
|
||
f.write(struct.pack("<I", len(payload))); f.write(payload)
|
||
|
||
envelope_size = Path(out_path).stat().st_size
|
||
dur = len(raw) / SR
|
||
kbps = envelope_size * 8 / dur / 1000
|
||
|
||
print(f"input audio: {dur:.2f} s mono @ {sr_in} Hz "
|
||
f"({raw_bytes:>10,} B as int16 PCM)")
|
||
print(f"Mimi tokens: {q} codebooks × {t} frames = {q*t} codes")
|
||
print(f"WAI envelope: {envelope_size:>10,} B "
|
||
f"({raw_bytes/envelope_size:6.1f}× compressed, "
|
||
f"{kbps:5.1f} kbps)")
|
||
print(f"model on the wire: 0 (Mimi weights stay at the sink)")
|
||
print(f"capability declared:{manifest['model_requirement']['capability']:>30}")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
p = argparse.ArgumentParser()
|
||
p.add_argument("input")
|
||
p.add_argument("-o", "--output", required=True)
|
||
a = p.parse_args()
|
||
encode(a.input, a.output)
|