[audio.cpp] New audio.cpp VoxCPM2 benchmark results on the RTX 5090 --- 0.140 RTF or 6.68x realtime.

[0xShug0]

Hi @a710128,

Here are the test cases and reproduction instructions:

1. Clone and build audio.cpp
   scripts/build_linux.sh --backend cuda --build-dir build/debug --build-type Debug --target audiocpp_cli --jobs 16

2. Download the GGUF models from https://huggingface.co/audio-cpp/audio.cpp-gguf

3. BF16 GGUF test:

python tools/audiocpp_cli/run_audiocpp_cli_path_tests.py \  
   --audiocpp-cli-bin build/debug/bin/audiocpp_cli \  
   --cases voxcpm2_clone_rtf_realistic_cases.json \
  --only voxcpm2_clone_rtf_realistic_interleaved \
  --backend cuda \
   --model-path /path/to/voxcpm2-bf16.gguf \
  --out-root build/logs/voxcpm2_clone_rtf_realistic_path_test

4. Check the results

python3 - <<'PY'
import re, wave
from pathlib import Path

case = Path("build/logs/voxcpm2_clone_rtf_realistic_path_test/voxcpm2_clone_rtf_realistic_interleaved")
timing = {}
order = []

for line in (case / "stdout.log").read_text().splitlines():
    m = re.match(r"\[TIMING\] request\.(.+)\.wall_ms ([0-9.]+)", line)
    if m:
        rid = m.group(1)
        timing[rid] = float(m.group(2))
        order.append(rid)

rows = []
for rid in order:
    if rid.startswith("warmup"):
        continue
    wav = case / "outputs" / f"{rid}.wav"
    with wave.open(str(wav), "rb") as w:
        audio_sec = w.getnframes() / w.getframerate()
    wall_ms = timing[rid]
    rtf = (wall_ms / 1000.0) / audio_sec
    rows.append((rid, audio_sec, wall_ms, rtf, 1.0 / rtf))

print("| request | audio sec | wall ms | RTF | x realtime |")
print("|---|---:|---:|---:|---:|")
for rid, audio_sec, wall_ms, rtf, speed in rows:
    print(f"| {rid} | {audio_sec:.2f} | {wall_ms:.1f} | {rtf:.3f} | {speed:.2f}x |")

avg_rtf = sum(r[3] for r in rows) / len(rows)
weighted_rtf = sum(r[2] for r in rows) / 1000.0 / sum(r[1] for r in rows)

print()
print(f"Average RTF: {avg_rtf:.3f}, x realtime: {1.0 / avg_rtf:.2f}x")
print(f"Weighted RTF: {weighted_rtf:.3f}, x realtime: {1.0 / weighted_rtf:.2f}x")
PY

BF 16

request	audio sec	wall ms	RTF	x realtime
short_r1	3.84	634.6	0.165	6.05x
medium_r1	9.12	1344.2	0.147	6.78x
long_r1	22.24	3199.9	0.144	6.95x
short_r2	3.84	756.4	0.197	5.08x
medium_r2	9.12	1438.5	0.158	6.34x
long_r2	22.24	3167.8	0.142	7.02x

Average RTF: 0.159, x realtime: 6.29x
Weighted RTF: 0.150, x realtime: 6.68x

Q8:

request	audio sec	wall ms	RTF	x realtime
short_r1	3.84	609.0	0.159	6.31x
medium_r1	8.80	1276.8	0.145	6.89x
long_r1	22.72	3236.3	0.142	7.02x
short_r2	3.84	751.9	0.196	5.11x
medium_r2	8.80	1401.3	0.159	6.28x
long_r2	22.72	3212.0	0.141	7.07x

Average RTF: 0.157, x realtime: 6.37x
Weighted RTF: 0.148, x realtime: 6.74x

voxcpm2_clone_rtf_realistic_cases.json

[0xShug0]

Forgot to mention: The reference voice is VoxCPM/examples/reference_speaker.wav. The file path is [audio.cpp repo root]/reference/VoxCPM/examples/reference_speaker.wav.