## Prerequisites

- Client ID and Client Secret from AudioShake ( [request access](mailto:info@audioshake.ai))
- A `.crypt` model file provided by AudioShake
- Platform requirements met — see [System Requirements](https://developer.audioshake.ai/sdk/requirements)

## File-based separation

Use `SourceSeparationTask` for file-to-file processing:

```
#include "AudioShakeSDK.h"

AudioFileReader* input = new AudioFileReader("input.wav");
WAVOutput* output = new WAVOutput("./output/");

SourceSeparationTask task(
    "your_client_id",
    "your_client_secret",
    input,
    output,
    "path/to/model.crypt"
);

void onProgress(SourceSeparationTask* task, double progress, void* data) {
    printf("Progress: %.0f%%\n", progress * 100);
}

if (!task.run(onProgress)) {
    fprintf(stderr, "Error: %s\n", task.getErrorMessage());
}
```

## Streaming separation

Use `RingBufferInput` and `RingBufferOutput` for real-time or streaming use cases:

```
RingBufferInput* input = new RingBufferInput(4096, true, 44100);
RingBufferOutput* output = new RingBufferOutput(4096, false);

SourceSeparationTask task(
    "your_client_id",
    "your_client_secret",
    input,
    output,
    "path/to/model.crypt"
);
```

Write audio frames into the input ring buffer, then read separated stems from the output ring buffer as processing proceeds.

## Choosing a chunk size

Chunk size controls the trade-off between latency and throughput. Durations are approximate and vary by model.

| Flag | Approx. duration | Use when |
| --- | --- | --- |
| `chunkNormal` (default) | ~3 sec | Batch or file processing |
| `chunk2X` | ~1.5 sec | Moderate latency |
| `chunk4X` | ~0.75 sec | Low latency |
| `chunk8X` | ~375 ms | Real-time |
| `chunk16X` | ~185 ms | Ultra low latency |
| `chunk32X` | ~92 ms | Ultra low latency |
| `chunk64X` | ~46 ms | Lowest latency |

Smaller chunks reduce latency but increase CPU/GPU overhead per unit of audio.