> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-coustics.com/llms.txt
> Use this file to discover all available pages before exploring further.
# C voice activity detection
> Process a dedicated VAD model and read its probability and filtered speech decision.
All functions on this page are from C SDK **0.25.0**, core **0.25.0**. Include `aic.h` and link the matching native library. Start with the [C integration guide](/reference/sdk/language-bindings/c).
The [pointer and error contract](/reference/sdk/api/c/errors-and-types#pointer-contract) applies to every signature below. Error names in prose omit the `AIC_ERROR_CODE_` prefix. There are no default C arguments; optional pointers are stated explicitly.
## AicVadParameter
All defaults are model-specific. Read the current values through `aic_vad_context_get_parameter` or `aic_energy_vad_context_get_parameter` after creation.
| Constant | Value | Units and constraints |
| - | - | - |
| `AIC_VAD_PARAMETER_SPEECH_HOLD_DURATION` | 0 | Seconds, from zero to the model/window-dependent maximum. Holds the speech decision through short gaps; internal decisions are quantized to model frames. |
| `AIC_VAD_PARAMETER_SENSITIVITY` | 1 | Meaning depends on the detector. On a dedicated VAD model, a probability threshold, 0–1 inclusive, where a higher threshold requires stronger evidence of speech. On an energy VAD (`AicEnergyVadContext`), the energy threshold of the enhanced signal, computed as `10 ^ (-sensitivity)` over 1–15 inclusive, where higher values detect quieter speech. |
| `AIC_VAD_PARAMETER_MINIMUM_SPEECH_DURATION` | 2 | Seconds, 0–1 inclusive. Requires sustained speech before the decision becomes true. |
Use a dedicated [ai-coustics VAD model](/models/voice-activity-detection/vad) and mono float32 audio. An enhancement model is not accepted by `aic_vad_create`; to detect speech with an enhancement model, create an energy VAD handle with [`aic_energy_vad_context_create`](/reference/sdk/api/c/vad#aic_energy_vad_context_create) instead.
## aic\_vad\_create
```c theme={null}
enum AicErrorCode aic_vad_create(struct AicVad **vad,
const struct AicModel *model,
const char *license_key,
const struct AicOtelConfig *otel_config);
```
`vad` is a writable output slot. `model` must be a VAD model; `license_key` is a null-terminated SDK key or supported bearer token. `otel_config` may be `NULL` to use environment defaults, or point to [AicOtelConfig](/reference/sdk/api/c/models-and-config#AicOtelConfig). The function copies configuration strings and retains model data internally. On success, destroy the owned VAD handle separately. Errors include `NULL_POINTER`, `MODEL_TYPE_UNSUPPORTED`, license errors and `INTERNAL_ERROR`. Construction allocates resources and may start session work; keep it outside an audio callback.
## aic\_vad\_destroy
```c theme={null}
void aic_vad_destroy(struct AicVad *vad);
```
Releases `vad` and its model reference and ends its associated session. `NULL` is a no-op. Stop processing and other calls that use this handle first. Context handles are independent and must be destroyed separately. Teardown can block; keep it outside an audio callback. No return value.
## aic\_vad\_initialize
```c theme={null}
enum AicErrorCode aic_vad_initialize(struct AicVad *vad,
uint32_t sample_rate,
size_t block_size,
bool variable_block_size);
```
Configures `vad` for `sample_rate` in Hz (8,000–192,000 inclusive), a positive `block_size` in mono samples and `variable_block_size`. With `false`, every call must supply exactly `block_size`; with `true`, that value is the maximum per call. Reinitialization resets processing state. Returns `SUCCESS`, `NULL_POINTER` or `AUDIO_CONFIG_UNSUPPORTED`. A failed initialization leaves the object uninitialized; initialize successfully before processing again. Allocates internal buffers; do not call concurrently with processing.
## aic\_vad\_process
```c theme={null}
enum AicErrorCode aic_vad_process(struct AicVad *vad,
const float *audio_ptr,
size_t audio_len);
```
`audio_ptr` points to `audio_len` readable contiguous mono float32 samples. The binding copies them to its internal buffer; the caller's input is unchanged. Processing updates the context's latest decision and probability. Returns `SUCCESS`, `NULL_POINTER`, `NOT_INITIALIZED`, `AUDIO_CONFIG_MISMATCH` or `PROCESSING_NOT_ALLOWED` (and internal failures). Read a fresh decision only after a successful call. Do not process the same object concurrently.
## aic\_vad\_terminate\_session
```c theme={null}
enum AicErrorCode aic_vad_terminate_session(struct AicVad *vad);
```
Requests termination of the session associated with `vad`. Returns `SUCCESS` after requesting it, or `NULL_POINTER`. Termination is handled asynchronously; a successful return does not prove remote acknowledgment or immediate processing denial. For credentials with no telemetry session, this is a no-op. It does not destroy the handle. Stop application processing explicitly and destroy the object when finished. On native targets, terminating the final active session may wait for the shared telemetry tasks to finish. Call it outside the audio callback.
## aic\_vad\_context\_create
```c theme={null}
enum AicErrorCode aic_vad_context_create(struct AicVadContext **context, const struct AicVad *vad);
```
`context` is a writable output slot; `vad` is the source object. Returns a separately owned context with shared control state. It can outlive the source object, but no new audio is processed after the source is destroyed. Returns `SUCCESS` or `NULL_POINTER` (or `INTERNAL_ERROR` if the VAD context is unavailable). Allocate outside the callback and destroy the context separately.
## aic\_vad\_context\_destroy
```c theme={null}
void aic_vad_context_destroy(struct AicVadContext *context);
```
Releases `context`; `NULL` is a no-op. Stop users of this context handle first. This does not destroy the processing object. No return value.
## aic\_vad\_context\_reset
```c theme={null}
enum AicErrorCode aic_vad_context_reset(const struct AicVadContext *context);
```
Requests processing-state reset for the next call and immediately clears the published speech decision to false and raw probability to zero. Audio configuration and parameter values remain in place. `context` must be valid. Returns `SUCCESS` or `NULL_POINTER`. The shared control operation can run alongside processing; coordinate a stream boundary if you need a precise reset position.
## aic\_vad\_context\_set\_parameter
```c theme={null}
enum AicErrorCode aic_vad_context_set_parameter(const struct AicVadContext *context,
enum AicVadParameter parameter,
float value);
```
Sets `parameter` to `value` through shared control state. Use only the declared [AicVadParameter](/reference/sdk/api/c/vad#AicVadParameter) values and documented ranges. Returns `SUCCESS`, `NULL_POINTER` or `PARAMETER_OUT_OF_RANGE`; NaN is out of range. Updates may run alongside processing. No ownership transfer.
## aic\_vad\_context\_get\_parameter
```c theme={null}
enum AicErrorCode aic_vad_context_get_parameter(const struct AicVadContext *context,
enum AicVadParameter parameter,
float *value);
```
Writes the current `parameter` value to the non-null writable `value` pointer. Use a declared enum value. Returns `SUCCESS` or `NULL_POINTER`. Shared control reads can run alongside processing. VAD duration values are stored in seconds; model-frame quantization affects decisions.
## aic\_vad\_context\_get\_prediction\_delay
```c theme={null}
enum AicErrorCode aic_vad_context_get_prediction_delay(const struct AicVadContext *context,
size_t *delay);
```
Writes the current prediction delay to `delay`, measured in samples at the initialized host rate. Query after successful initialization and again if you reinitialize. Both `context` and writable `delay` must be non-null. Returns `SUCCESS` or `NULL_POINTER`. This shared metadata read can run alongside processing; it is not elapsed CPU time.
## aic\_vad\_context\_update\_bearer\_token
```c theme={null}
enum AicErrorCode aic_vad_context_update_bearer_token(const struct AicVadContext *context,
const char *token);
```
Updates credentials through `context` using null-terminated UTF-8 `token`. Both the original credential and replacement must use the supported JWT form. Returns `SUCCESS`, `NULL_POINTER`, `LICENSE_FORMAT_INVALID`, other license errors or `TOKEN_UPDATE_UNSUPPORTED`. Success updates local credential state; it does not prove backend acceptance. This is a control-thread operation, not an audio-callback operation. See [authentication](/models/get-started/authenticate-apps).
## aic\_vad\_context\_is\_speech\_detected
```c theme={null}
enum AicErrorCode aic_vad_context_is_speech_detected(const struct AicVadContext *context,
bool *value);
```
Writes the latest filtered speech decision into the writable `value` pointer. Returns `SUCCESS` or `NULL_POINTER`. This does not process audio; call `aic_vad_process` first and check its result. It combines model probability, sensitivity, minimum speech duration and hold behavior. Safe to read alongside processing while the context remains alive.
## aic\_vad\_context\_get\_raw\_vad\_probability
```c theme={null}
enum AicErrorCode aic_vad_context_get_raw_vad_probability(const struct AicVadContext *context,
float *value);
```
Writes the latest model speech probability into the writable `value` pointer. Returns `SUCCESS` or `NULL_POINTER`. This is the probability before duration filtering, not the final speech decision. It does not process audio. Safe to read alongside processing while the context remains alive.
## aic\_energy\_vad\_context\_create
```c theme={null}
enum AicErrorCode aic_energy_vad_context_create(struct AicEnergyVadContext **context,
const struct AicProcessor *processor);
```
`context` is a writable output slot; `processor` is an enhancement processor. Returns a separately owned handle to that processor's energy VAD, which detects speech from the enhanced signal before output mixing and runs no separate model. Detection advances with `aic_processor_process`; there is no separate process call. Every handle from one processor refers to the same energy VAD.
Creating a handle keeps enhancement inference active even when the processor is bypassed or its enhancement level is zero, and it stays active for the processor's lifetime after all handles are destroyed. The handle can outlive the processor, but publishes no new predictions once the processor is destroyed. Returns `SUCCESS`, `NULL_POINTER` or `MODEL_TYPE_UNSUPPORTED` when the processor does not support energy VAD. Allocates memory; keep it outside an audio callback and do not use or destroy the processor from another thread during the call.
## aic\_energy\_vad\_context\_destroy
```c theme={null}
void aic_energy_vad_context_destroy(struct AicEnergyVadContext *context);
```
Releases `context`; `NULL` is a no-op. This does not destroy the backing processor or disable its energy VAD; call `aic_processor_destroy` separately. Stop users of this handle first. Deallocates memory, so keep it outside an audio callback. No return value.
## aic\_energy\_vad\_context\_reset
```c theme={null}
enum AicErrorCode aic_energy_vad_context_reset(const struct AicEnergyVadContext *context);
```
Clears the energy VAD state and the published speech decision. Use it when the stream is interrupted or when seeking. Parameter values are retained and the backing processor is not reset; `aic_processor_context_reset` also resets the energy VAD. `context` must be valid. Returns `SUCCESS` or `NULL_POINTER`. Real-time safe and callable from any thread.
## aic\_energy\_vad\_context\_set\_parameter
```c theme={null}
enum AicErrorCode aic_energy_vad_context_set_parameter(const struct AicEnergyVadContext *context,
enum AicVadParameter parameter,
float value);
```
Sets `parameter` to `value`. Use the declared [AicVadParameter](/reference/sdk/api/c/vad#AicVadParameter) values; `AIC_VAD_PARAMETER_SENSITIVITY` takes the energy range, 1.0 to 15.0, on this handle. Returns `SUCCESS`, `NULL_POINTER` or `PARAMETER_OUT_OF_RANGE`. All parameters can be changed during processing, from any thread.
## aic\_energy\_vad\_context\_get\_parameter
```c theme={null}
enum AicErrorCode aic_energy_vad_context_get_parameter(const struct AicEnergyVadContext *context,
enum AicVadParameter parameter,
float *value);
```
Writes the current `parameter` value to the non-null writable `value` pointer. Use a declared enum value. Returns `SUCCESS` or `NULL_POINTER`. Real-time safe and callable from any thread.
## aic\_energy\_vad\_context\_is\_speech\_detected
```c theme={null}
enum AicErrorCode aic_energy_vad_context_is_speech_detected(const struct AicEnergyVadContext *context,
bool *value);
```
Writes the energy VAD's prediction into the writable `value` pointer. It is false before audio is processed and after a reset, and stops updating if the backing processor stops being processed. The decision lags its input by `aic_energy_vad_context_get_prediction_delay` samples; speech hold and minimum speech duration affect its timing further. Returns `SUCCESS` or `NULL_POINTER`. Real-time safe and callable from any thread.
## aic\_energy\_vad\_context\_get\_prediction\_delay
```c theme={null}
enum AicErrorCode aic_energy_vad_context_get_prediction_delay(const struct AicEnergyVadContext *context,
size_t *delay);
```
Writes the prediction delay in samples to `delay`, covering input reblocking, STFT and enhancement model processing. It equals `aic_processor_context_get_audio_delay` for the backing processor; energy detection adds no audio delay. Speech hold and minimum speech duration affect decision timing and are not included.
Before initialization the value uses the processor's optimal block size at its native sample rate. After initialization it is expressed in samples at the rate given to `aic_processor_initialize` and includes input-buffering latency for the configured block size, which a nonoptimal or variable block size increases. Convert with `delay_ms = (delay_samples * 1000) / sample_rate`. Both pointers must be non-null. Returns `SUCCESS` or `NULL_POINTER`. Real-time safe and callable from any thread.
## Related
[C API index](/reference/sdk/api/c/index) · [C examples](/reference/sdk/examples#c) · [Compatibility](/reference/sdk/compatibility-matrix)
## Example: read a fresh decision
Use an initialized VAD and its context. The input is read-only; the decision is written only after successful processing.
```c theme={null}
#include "aic.h"
AicErrorCode detect_block(AicVad *vad, AicVadContext *context,
const float *audio, size_t count, bool *speech) {
AicErrorCode error = aic_vad_process(vad, audio, count);
if (error != AIC_ERROR_CODE_SUCCESS) return error;
return aic_vad_context_is_speech_detected(context, speech);
}
```
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.