WAV Audio Analysis Skill

Description

Analyze WAV audio files to debug audio generation pipelines. Provides statistical analysis, format validation, and quality metrics for diagnosing issues with generated speech.

Triggers: wav, audio, waveform, samples, amplitude, audio analysis, sound quality, audio debug

Analysis Capabilities

Basic Statistics

• Sample count and duration
• Min/max amplitude
• Standard deviation (expected ~3000-8000 for speech)
• Near-silent sample percentage

Quality Indicators

• Zero crossing rate (speech typically 50-200 per 1000 samples)
• Clipping detection (samples at ±32767)
• NaN/Inf detection (if processing raw floats)
• DC offset analysis

Format Validation

• Sample rate verification (24kHz for Qwen3-Omni TTS)
• Bit depth check
• Channel count
• RIFF header validation

Usage

To analyze a WAV file, provide the path and I'll run comprehensive diagnostics:

import numpy as np

with open("audio.wav", "rb") as f:
    header = f.read(44)
    data = f.read()

samples = np.frombuffer(data, dtype=np.int16)
print(f"Samples: {len(samples)}")
print(f"Duration: {len(samples)/24000:.2f} sec")
print(f"Min/Max: {samples.min()} / {samples.max()}")
print(f"Std dev: {np.std(samples):.1f}")

# Quality check
near_silent = np.sum(np.abs(samples) < 100)
print(f"Near-silent: {100*near_silent/len(samples):.1f}%")

# Zero crossings (voice activity indicator)
if len(samples) > 1000:
    zc = np.sum(np.diff(np.sign(samples[:1000])) != 0)
    print(f"Zero crossings (first 1000): {zc}")

Typical Values for Good Speech Audio

Metric	Expected Range	Meaning
Std dev	3000-8000	Audio energy level
Near-silent	<5%	Minimal silent padding
Zero crossings	50-200/1000	Voice frequency activity
Min/Max	±20000-32000	Healthy amplitude range

Common Issues

99% Near-Silent

• Cause: NaN values converted to zeros
• Fix: Check for numerical overflow in pipeline

Low Std Dev (<1000)

• Cause: Values too quiet before output normalization
• Fix: Check gain stages, ensure proper scaling

Constant Value Runs

• Cause: Chunked processing with context overlap issues
• Fix: Verify chunk stitching logic

Clipping (values at ±32767)

• Cause: Overflow or missing tanh/clamp
• Fix: Add output clamping before int16 conversion