> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/openai/whisper/llms.txt
> Use this file to discover all available pages before exploring further.

# load_audio

> Load an audio file and convert to mono waveform with resampling

## Function Signature

```python theme={null}
load_audio(file: str, sr: int = SAMPLE_RATE) -> np.ndarray
```

Opens an audio file and reads it as a mono waveform, resampling as necessary. This function uses FFmpeg to decode audio while down-mixing and resampling.

## Parameters

<ParamField path="file" type="str" required>
  The path to the audio file to open. Supports any audio format that FFmpeg can decode.
</ParamField>

<ParamField path="sr" type="int" default="16000">
  The sample rate to resample the audio if necessary. Defaults to `SAMPLE_RATE` (16000 Hz).
</ParamField>

## Returns

<ResponseField name="waveform" type="np.ndarray">
  A NumPy array containing the audio waveform in `float32` dtype. Values are normalized to the range \[-1.0, 1.0].
</ResponseField>

## Example

```python theme={null}
import whisper
from whisper.audio import load_audio

# Load audio file at default 16kHz sample rate
audio = load_audio("speech.mp3")

# Load audio at custom sample rate
audio = load_audio("speech.wav", sr=22050)

# Audio is returned as normalized float32 array
print(audio.dtype)  # float32
print(audio.min(), audio.max())  # Values in range [-1.0, 1.0]
```

## Implementation Details

### FFmpeg Dependency

This function requires the FFmpeg CLI to be available in your system PATH. It launches a subprocess with the following operations:

* Decodes the input audio file
* Down-mixes to mono (`-ac 1`)
* Resamples to the specified sample rate
* Outputs as 16-bit PCM (`-f s16le`)

### Normalization

The raw 16-bit PCM output is converted to float32 and normalized by dividing by 32768.0, mapping the integer range \[-32768, 32767] to the float range \[-1.0, 1.0].

## Audio Constants

The default sample rate and other audio constants used in Whisper:

```python theme={null}
SAMPLE_RATE = 16000  # 16 kHz sample rate
N_FFT = 400          # FFT window size
HOP_LENGTH = 160     # Number of samples between STFT columns
CHUNK_LENGTH = 30    # 30-second chunks
N_SAMPLES = 480000   # Samples in a 30-second chunk (CHUNK_LENGTH * SAMPLE_RATE)
```

## Error Handling

```python theme={null}
try:
    audio = load_audio("invalid_file.mp3")
except RuntimeError as e:
    print(f"Failed to load audio: {e}")
```

Raises `RuntimeError` if FFmpeg fails to decode the audio file, with the stderr output from FFmpeg included in the error message.