AudioFrequencyConverter silently turns floating-point WAV input into silence

[Garyouki]

Describe the bug

AudioFrequencyConverter reads floating-point WAV data successfully, but casts the shifted samples directly to int16 without scaling. A valid floating-point signal whose samples have magnitude below 1 is therefore written as all-zero audio. Conversion returns successfully, so the caller receives a silent artifact rather than an error. Even shift_value=0 loses the signal.

The cast is in pyrit/converter/audio_frequency_converter.py, immediately before wavfile.write. Existing success tests exercise int16 data. This differs from #2356, which fixed stereo broadcasting.

Steps/Code to Reproduce

Run from an environment with PyRIT installed:

import asyncio
import tempfile
from pathlib import Path

import numpy as np
from scipy.io import wavfile

from pyrit.converter import AudioFrequencyConverter
from pyrit.memory import CentralMemory, SQLiteMemory


async def main_async():
    with tempfile.TemporaryDirectory() as directory:
        memory = SQLiteMemory(db_path=':memory:')
        memory.results_path = directory
        CentralMemory.set_memory_instance(memory)
        try:
            source = Path(directory) / 'float.wav'
            samples = np.array([0.0, 0.5, -0.5, 0.25], dtype=np.float32)
            wavfile.write(source, 8000, samples)
            result = await AudioFrequencyConverter(shift_value=0).convert_async(prompt=str(source))
            rate, output = wavfile.read(result.output_text)
            print('frequency:', output.dtype, output.tolist())
        finally:
            memory.dispose_engine()


asyncio.run(main_async())

Expected Results

A zero frequency shift preserves the input waveform and amplitude. Floating-point samples should either retain their representation or be correctly scaled if converted to PCM; silently truncating them is incorrect.

Actual Results

frequency: int16 [0, 0, 0, 0]

The input is float32 [0.0, 0.5, -0.5, 0.25].

Versions

Linux, Python 3.12.3, PyRIT 1.2.0.dev0 at 8934a3c4, NumPy 2.5.3, SciPy 1.18.1. Reproduced locally through the public convert_async methods with real WAV files; no model or external service is used.