2 * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
11 #ifndef WEBRTC_MODULES_AUDIO_PROCESSING_INCLUDE_AUDIO_PROCESSING_H_
12 #define WEBRTC_MODULES_AUDIO_PROCESSING_INCLUDE_AUDIO_PROCESSING_H_
14 #include <stddef.h> // size_t
16 #include "webrtc/common.h"
17 #include "webrtc/modules/interface/module.h"
18 #include "webrtc/typedefs.h"
25 class EchoCancellation;
26 class EchoControlMobile;
30 class NoiseSuppression;
33 // Use to enable the delay correction feature. This now engages an extended
34 // filter mode in the AEC, along with robustness measures around the reported
35 // system delays. It comes with a significant increase in AEC complexity, but is
36 // much more robust to unreliable reported delays.
38 // Detailed changes to the algorithm:
39 // - The filter length is changed from 48 to 128 ms. This comes with tuning of
40 // several parameters: i) filter adaptation stepsize and error threshold;
41 // ii) non-linear processing smoothing and overdrive.
42 // - Option to ignore the reported delays on platforms which we deem
43 // sufficiently unreliable. See WEBRTC_UNTRUSTED_DELAY in echo_cancellation.c.
44 // - Faster startup times by removing the excessive "startup phase" processing
45 // of reported delays.
46 // - Much more conservative adjustments to the far-end read pointer. We smooth
47 // the delay difference more heavily, and back off from the difference more.
48 // Adjustments force a readaptation of the filter, so they should be avoided
49 // except when really necessary.
50 struct DelayCorrection {
51 DelayCorrection() : enabled(false) {}
52 DelayCorrection(bool enabled) : enabled(enabled) {}
57 // The Audio Processing Module (APM) provides a collection of voice processing
58 // components designed for real-time communications software.
60 // APM operates on two audio streams on a frame-by-frame basis. Frames of the
61 // primary stream, on which all processing is applied, are passed to
62 // |ProcessStream()|. Frames of the reverse direction stream, which are used for
63 // analysis by some components, are passed to |AnalyzeReverseStream()|. On the
64 // client-side, this will typically be the near-end (capture) and far-end
65 // (render) streams, respectively. APM should be placed in the signal chain as
66 // close to the audio hardware abstraction layer (HAL) as possible.
68 // On the server-side, the reverse stream will normally not be used, with
69 // processing occurring on each incoming stream.
71 // Component interfaces follow a similar pattern and are accessed through
72 // corresponding getters in APM. All components are disabled at create-time,
73 // with default settings that are recommended for most situations. New settings
74 // can be applied without enabling a component. Enabling a component triggers
75 // memory allocation and initialization to allow it to start processing the
78 // Thread safety is provided with the following assumptions to reduce locking
80 // 1. The stream getters and setters are called from the same thread as
81 // ProcessStream(). More precisely, stream functions are never called
82 // concurrently with ProcessStream().
83 // 2. Parameter getters are never called concurrently with the corresponding
86 // APM accepts only 16-bit linear PCM audio data in frames of 10 ms. Multiple
87 // channels should be interleaved.
89 // Usage example, omitting error checking:
90 // AudioProcessing* apm = AudioProcessing::Create(0);
91 // apm->set_sample_rate_hz(32000); // Super-wideband processing.
93 // // Mono capture and stereo render.
94 // apm->set_num_channels(1, 1);
95 // apm->set_num_reverse_channels(2);
97 // apm->high_pass_filter()->Enable(true);
99 // apm->echo_cancellation()->enable_drift_compensation(false);
100 // apm->echo_cancellation()->Enable(true);
102 // apm->noise_reduction()->set_level(kHighSuppression);
103 // apm->noise_reduction()->Enable(true);
105 // apm->gain_control()->set_analog_level_limits(0, 255);
106 // apm->gain_control()->set_mode(kAdaptiveAnalog);
107 // apm->gain_control()->Enable(true);
109 // apm->voice_detection()->Enable(true);
111 // // Start a voice call...
113 // // ... Render frame arrives bound for the audio HAL ...
114 // apm->AnalyzeReverseStream(render_frame);
116 // // ... Capture frame arrives from the audio HAL ...
117 // // Call required set_stream_ functions.
118 // apm->set_stream_delay_ms(delay_ms);
119 // apm->gain_control()->set_stream_analog_level(analog_level);
121 // apm->ProcessStream(capture_frame);
123 // // Call required stream_ functions.
124 // analog_level = apm->gain_control()->stream_analog_level();
125 // has_voice = apm->stream_has_voice();
127 // // Repeate render and capture processing for the duration of the call...
128 // // Start a new call...
129 // apm->Initialize();
131 // // Close the application...
134 class AudioProcessing : public Module {
136 // Creates a APM instance, with identifier |id|. Use one instance for every
137 // primary audio stream requiring processing. On the client-side, this would
138 // typically be one instance for the near-end stream, and additional instances
139 // for each far-end stream which requires processing. On the server-side,
140 // this would typically be one instance for every incoming stream.
141 static AudioProcessing* Create(int id);
142 virtual ~AudioProcessing() {}
144 // Initializes internal states, while retaining all user settings. This
145 // should be called before beginning to process a new audio stream. However,
146 // it is not necessary to call before processing the first stream after
149 // set_sample_rate_hz(), set_num_channels() and set_num_reverse_channels()
150 // will trigger a full initialization if the settings are changed from their
151 // existing values. Otherwise they are no-ops.
152 virtual int Initialize() = 0;
154 // Pass down additional options which don't have explicit setters. This
155 // ensures the options are applied immediately.
156 virtual void SetExtraOptions(const Config& config) = 0;
158 // Sets the sample |rate| in Hz for both the primary and reverse audio
159 // streams. 8000, 16000 or 32000 Hz are permitted.
160 virtual int set_sample_rate_hz(int rate) = 0;
161 virtual int sample_rate_hz() const = 0;
163 // Sets the number of channels for the primary audio stream. Input frames must
164 // contain a number of channels given by |input_channels|, while output frames
165 // will be returned with number of channels given by |output_channels|.
166 virtual int set_num_channels(int input_channels, int output_channels) = 0;
167 virtual int num_input_channels() const = 0;
168 virtual int num_output_channels() const = 0;
170 // Sets the number of channels for the reverse audio stream. Input frames must
171 // contain a number of channels given by |channels|.
172 virtual int set_num_reverse_channels(int channels) = 0;
173 virtual int num_reverse_channels() const = 0;
175 // Processes a 10 ms |frame| of the primary audio stream. On the client-side,
176 // this is the near-end (or captured) audio.
178 // If needed for enabled functionality, any function with the set_stream_ tag
179 // must be called prior to processing the current frame. Any getter function
180 // with the stream_ tag which is needed should be called after processing.
182 // The |sample_rate_hz_|, |num_channels_|, and |samples_per_channel_|
183 // members of |frame| must be valid, and correspond to settings supplied
185 virtual int ProcessStream(AudioFrame* frame) = 0;
187 // Analyzes a 10 ms |frame| of the reverse direction audio stream. The frame
188 // will not be modified. On the client-side, this is the far-end (or to be
191 // It is only necessary to provide this if echo processing is enabled, as the
192 // reverse stream forms the echo reference signal. It is recommended, but not
193 // necessary, to provide if gain control is enabled. On the server-side this
194 // typically will not be used. If you're not sure what to pass in here,
195 // chances are you don't need to use it.
197 // The |sample_rate_hz_|, |num_channels_|, and |samples_per_channel_|
198 // members of |frame| must be valid.
200 // TODO(ajm): add const to input; requires an implementation fix.
201 virtual int AnalyzeReverseStream(AudioFrame* frame) = 0;
203 // This must be called if and only if echo processing is enabled.
205 // Sets the |delay| in ms between AnalyzeReverseStream() receiving a far-end
206 // frame and ProcessStream() receiving a near-end frame containing the
207 // corresponding echo. On the client-side this can be expressed as
208 // delay = (t_render - t_analyze) + (t_process - t_capture)
210 // - t_analyze is the time a frame is passed to AnalyzeReverseStream() and
211 // t_render is the time the first sample of the same frame is rendered by
212 // the audio hardware.
213 // - t_capture is the time the first sample of a frame is captured by the
214 // audio hardware and t_pull is the time the same frame is passed to
216 virtual int set_stream_delay_ms(int delay) = 0;
217 virtual int stream_delay_ms() const = 0;
219 // Sets a delay |offset| in ms to add to the values passed in through
220 // set_stream_delay_ms(). May be positive or negative.
222 // Note that this could cause an otherwise valid value passed to
223 // set_stream_delay_ms() to return an error.
224 virtual void set_delay_offset_ms(int offset) = 0;
225 virtual int delay_offset_ms() const = 0;
227 // Starts recording debugging information to a file specified by |filename|,
228 // a NULL-terminated string. If there is an ongoing recording, the old file
229 // will be closed, and recording will continue in the newly specified file.
230 // An already existing file will be overwritten without warning.
231 static const size_t kMaxFilenameSize = 1024;
232 virtual int StartDebugRecording(const char filename[kMaxFilenameSize]) = 0;
234 // Stops recording debugging information, and closes the file. Recording
235 // cannot be resumed in the same file (without overwriting it).
236 virtual int StopDebugRecording() = 0;
238 // These provide access to the component interfaces and should never return
239 // NULL. The pointers will be valid for the lifetime of the APM instance.
240 // The memory for these objects is entirely managed internally.
241 virtual EchoCancellation* echo_cancellation() const = 0;
242 virtual EchoControlMobile* echo_control_mobile() const = 0;
243 virtual GainControl* gain_control() const = 0;
244 virtual HighPassFilter* high_pass_filter() const = 0;
245 virtual LevelEstimator* level_estimator() const = 0;
246 virtual NoiseSuppression* noise_suppression() const = 0;
247 virtual VoiceDetection* voice_detection() const = 0;
250 int instant; // Instantaneous value.
251 int average; // Long-term average.
252 int maximum; // Long-term maximum.
253 int minimum; // Long-term minimum.
259 kUnspecifiedError = -1,
260 kCreationFailedError = -2,
261 kUnsupportedComponentError = -3,
262 kUnsupportedFunctionError = -4,
263 kNullPointerError = -5,
264 kBadParameterError = -6,
265 kBadSampleRateError = -7,
266 kBadDataLengthError = -8,
267 kBadNumberChannelsError = -9,
269 kStreamParameterNotSetError = -11,
270 kNotEnabledError = -12,
272 // Warnings are non-fatal.
273 // This results when a set_stream_ parameter is out of range. Processing
274 // will continue, but the parameter may have been truncated.
275 kBadStreamParameterWarning = -13
278 // Inherited from Module.
279 virtual int32_t TimeUntilNextProcess() OVERRIDE;
280 virtual int32_t Process() OVERRIDE;
283 // The acoustic echo cancellation (AEC) component provides better performance
284 // than AECM but also requires more processing power and is dependent on delay
285 // stability and reporting accuracy. As such it is well-suited and recommended
286 // for PC and IP phone applications.
288 // Not recommended to be enabled on the server-side.
289 class EchoCancellation {
291 // EchoCancellation and EchoControlMobile may not be enabled simultaneously.
292 // Enabling one will disable the other.
293 virtual int Enable(bool enable) = 0;
294 virtual bool is_enabled() const = 0;
296 // Differences in clock speed on the primary and reverse streams can impact
297 // the AEC performance. On the client-side, this could be seen when different
298 // render and capture devices are used, particularly with webcams.
300 // This enables a compensation mechanism, and requires that
301 // |set_device_sample_rate_hz()| and |set_stream_drift_samples()| be called.
302 virtual int enable_drift_compensation(bool enable) = 0;
303 virtual bool is_drift_compensation_enabled() const = 0;
305 // Provides the sampling rate of the audio devices. It is assumed the render
306 // and capture devices use the same nominal sample rate. Required if and only
307 // if drift compensation is enabled.
308 virtual int set_device_sample_rate_hz(int rate) = 0;
309 virtual int device_sample_rate_hz() const = 0;
311 // Sets the difference between the number of samples rendered and captured by
312 // the audio devices since the last call to |ProcessStream()|. Must be called
313 // if drift compensation is enabled, prior to |ProcessStream()|.
314 virtual void set_stream_drift_samples(int drift) = 0;
315 virtual int stream_drift_samples() const = 0;
317 enum SuppressionLevel {
319 kModerateSuppression,
323 // Sets the aggressiveness of the suppressor. A higher level trades off
324 // double-talk performance for increased echo suppression.
325 virtual int set_suppression_level(SuppressionLevel level) = 0;
326 virtual SuppressionLevel suppression_level() const = 0;
328 // Returns false if the current frame almost certainly contains no echo
329 // and true if it _might_ contain echo.
330 virtual bool stream_has_echo() const = 0;
332 // Enables the computation of various echo metrics. These are obtained
333 // through |GetMetrics()|.
334 virtual int enable_metrics(bool enable) = 0;
335 virtual bool are_metrics_enabled() const = 0;
337 // Each statistic is reported in dB.
338 // P_far: Far-end (render) signal power.
339 // P_echo: Near-end (capture) echo signal power.
340 // P_out: Signal power at the output of the AEC.
341 // P_a: Internal signal power at the point before the AEC's non-linear
345 AudioProcessing::Statistic residual_echo_return_loss;
347 // ERL = 10log_10(P_far / P_echo)
348 AudioProcessing::Statistic echo_return_loss;
350 // ERLE = 10log_10(P_echo / P_out)
351 AudioProcessing::Statistic echo_return_loss_enhancement;
353 // (Pre non-linear processing suppression) A_NLP = 10log_10(P_echo / P_a)
354 AudioProcessing::Statistic a_nlp;
357 // TODO(ajm): discuss the metrics update period.
358 virtual int GetMetrics(Metrics* metrics) = 0;
360 // Enables computation and logging of delay values. Statistics are obtained
361 // through |GetDelayMetrics()|.
362 virtual int enable_delay_logging(bool enable) = 0;
363 virtual bool is_delay_logging_enabled() const = 0;
365 // The delay metrics consists of the delay |median| and the delay standard
366 // deviation |std|. The values are averaged over the time period since the
367 // last call to |GetDelayMetrics()|.
368 virtual int GetDelayMetrics(int* median, int* std) = 0;
370 // Returns a pointer to the low level AEC component. In case of multiple
371 // channels, the pointer to the first one is returned. A NULL pointer is
372 // returned when the AEC component is disabled or has not been initialized
374 virtual struct AecCore* aec_core() const = 0;
377 virtual ~EchoCancellation() {}
380 // The acoustic echo control for mobile (AECM) component is a low complexity
381 // robust option intended for use on mobile devices.
383 // Not recommended to be enabled on the server-side.
384 class EchoControlMobile {
386 // EchoCancellation and EchoControlMobile may not be enabled simultaneously.
387 // Enabling one will disable the other.
388 virtual int Enable(bool enable) = 0;
389 virtual bool is_enabled() const = 0;
391 // Recommended settings for particular audio routes. In general, the louder
392 // the echo is expected to be, the higher this value should be set. The
393 // preferred setting may vary from device to device.
395 kQuietEarpieceOrHeadset,
402 // Sets echo control appropriate for the audio routing |mode| on the device.
403 // It can and should be updated during a call if the audio routing changes.
404 virtual int set_routing_mode(RoutingMode mode) = 0;
405 virtual RoutingMode routing_mode() const = 0;
407 // Comfort noise replaces suppressed background noise to maintain a
408 // consistent signal level.
409 virtual int enable_comfort_noise(bool enable) = 0;
410 virtual bool is_comfort_noise_enabled() const = 0;
412 // A typical use case is to initialize the component with an echo path from a
413 // previous call. The echo path is retrieved using |GetEchoPath()|, typically
414 // at the end of a call. The data can then be stored for later use as an
415 // initializer before the next call, using |SetEchoPath()|.
417 // Controlling the echo path this way requires the data |size_bytes| to match
418 // the internal echo path size. This size can be acquired using
419 // |echo_path_size_bytes()|. |SetEchoPath()| causes an entire reset, worth
420 // noting if it is to be called during an ongoing call.
422 // It is possible that version incompatibilities may result in a stored echo
423 // path of the incorrect size. In this case, the stored path should be
425 virtual int SetEchoPath(const void* echo_path, size_t size_bytes) = 0;
426 virtual int GetEchoPath(void* echo_path, size_t size_bytes) const = 0;
428 // The returned path size is guaranteed not to change for the lifetime of
430 static size_t echo_path_size_bytes();
433 virtual ~EchoControlMobile() {}
436 // The automatic gain control (AGC) component brings the signal to an
437 // appropriate range. This is done by applying a digital gain directly and, in
438 // the analog mode, prescribing an analog gain to be applied at the audio HAL.
440 // Recommended to be enabled on the client-side.
443 virtual int Enable(bool enable) = 0;
444 virtual bool is_enabled() const = 0;
446 // When an analog mode is set, this must be called prior to |ProcessStream()|
447 // to pass the current analog level from the audio HAL. Must be within the
448 // range provided to |set_analog_level_limits()|.
449 virtual int set_stream_analog_level(int level) = 0;
451 // When an analog mode is set, this should be called after |ProcessStream()|
452 // to obtain the recommended new analog level for the audio HAL. It is the
453 // users responsibility to apply this level.
454 virtual int stream_analog_level() = 0;
457 // Adaptive mode intended for use if an analog volume control is available
458 // on the capture device. It will require the user to provide coupling
459 // between the OS mixer controls and AGC through the |stream_analog_level()|
462 // It consists of an analog gain prescription for the audio device and a
463 // digital compression stage.
466 // Adaptive mode intended for situations in which an analog volume control
467 // is unavailable. It operates in a similar fashion to the adaptive analog
468 // mode, but with scaling instead applied in the digital domain. As with
469 // the analog mode, it additionally uses a digital compression stage.
472 // Fixed mode which enables only the digital compression stage also used by
473 // the two adaptive modes.
475 // It is distinguished from the adaptive modes by considering only a
476 // short time-window of the input signal. It applies a fixed gain through
477 // most of the input level range, and compresses (gradually reduces gain
478 // with increasing level) the input signal at higher levels. This mode is
479 // preferred on embedded devices where the capture signal level is
480 // predictable, so that a known gain can be applied.
484 virtual int set_mode(Mode mode) = 0;
485 virtual Mode mode() const = 0;
487 // Sets the target peak |level| (or envelope) of the AGC in dBFs (decibels
488 // from digital full-scale). The convention is to use positive values. For
489 // instance, passing in a value of 3 corresponds to -3 dBFs, or a target
490 // level 3 dB below full-scale. Limited to [0, 31].
492 // TODO(ajm): use a negative value here instead, if/when VoE will similarly
493 // update its interface.
494 virtual int set_target_level_dbfs(int level) = 0;
495 virtual int target_level_dbfs() const = 0;
497 // Sets the maximum |gain| the digital compression stage may apply, in dB. A
498 // higher number corresponds to greater compression, while a value of 0 will
499 // leave the signal uncompressed. Limited to [0, 90].
500 virtual int set_compression_gain_db(int gain) = 0;
501 virtual int compression_gain_db() const = 0;
503 // When enabled, the compression stage will hard limit the signal to the
504 // target level. Otherwise, the signal will be compressed but not limited
505 // above the target level.
506 virtual int enable_limiter(bool enable) = 0;
507 virtual bool is_limiter_enabled() const = 0;
509 // Sets the |minimum| and |maximum| analog levels of the audio capture device.
510 // Must be set if and only if an analog mode is used. Limited to [0, 65535].
511 virtual int set_analog_level_limits(int minimum,
513 virtual int analog_level_minimum() const = 0;
514 virtual int analog_level_maximum() const = 0;
516 // Returns true if the AGC has detected a saturation event (period where the
517 // signal reaches digital full-scale) in the current frame and the analog
518 // level cannot be reduced.
520 // This could be used as an indicator to reduce or disable analog mic gain at
522 virtual bool stream_is_saturated() const = 0;
525 virtual ~GainControl() {}
528 // A filtering component which removes DC offset and low-frequency noise.
529 // Recommended to be enabled on the client-side.
530 class HighPassFilter {
532 virtual int Enable(bool enable) = 0;
533 virtual bool is_enabled() const = 0;
536 virtual ~HighPassFilter() {}
539 // An estimation component used to retrieve level metrics.
540 class LevelEstimator {
542 virtual int Enable(bool enable) = 0;
543 virtual bool is_enabled() const = 0;
545 // Returns the root mean square (RMS) level in dBFs (decibels from digital
546 // full-scale), or alternately dBov. It is computed over all primary stream
547 // frames since the last call to RMS(). The returned value is positive but
548 // should be interpreted as negative. It is constrained to [0, 127].
550 // The computation follows:
551 // http://tools.ietf.org/html/draft-ietf-avtext-client-to-mixer-audio-level-05
552 // with the intent that it can provide the RTP audio level indication.
554 // Frames passed to ProcessStream() with an |_energy| of zero are considered
555 // to have been muted. The RMS of the frame will be interpreted as -127.
556 virtual int RMS() = 0;
559 virtual ~LevelEstimator() {}
562 // The noise suppression (NS) component attempts to remove noise while
563 // retaining speech. Recommended to be enabled on the client-side.
565 // Recommended to be enabled on the client-side.
566 class NoiseSuppression {
568 virtual int Enable(bool enable) = 0;
569 virtual bool is_enabled() const = 0;
571 // Determines the aggressiveness of the suppression. Increasing the level
572 // will reduce the noise level at the expense of a higher speech distortion.
580 virtual int set_level(Level level) = 0;
581 virtual Level level() const = 0;
583 // Returns the internally computed prior speech probability of current frame
584 // averaged over output channels. This is not supported in fixed point, for
585 // which |kUnsupportedFunctionError| is returned.
586 virtual float speech_probability() const = 0;
589 virtual ~NoiseSuppression() {}
592 // The voice activity detection (VAD) component analyzes the stream to
593 // determine if voice is present. A facility is also provided to pass in an
594 // external VAD decision.
596 // In addition to |stream_has_voice()| the VAD decision is provided through the
597 // |AudioFrame| passed to |ProcessStream()|. The |vad_activity_| member will be
598 // modified to reflect the current decision.
599 class VoiceDetection {
601 virtual int Enable(bool enable) = 0;
602 virtual bool is_enabled() const = 0;
604 // Returns true if voice is detected in the current frame. Should be called
605 // after |ProcessStream()|.
606 virtual bool stream_has_voice() const = 0;
608 // Some of the APM functionality requires a VAD decision. In the case that
609 // a decision is externally available for the current frame, it can be passed
610 // in here, before |ProcessStream()| is called.
612 // VoiceDetection does _not_ need to be enabled to use this. If it happens to
613 // be enabled, detection will be skipped for any frame in which an external
614 // VAD decision is provided.
615 virtual int set_stream_has_voice(bool has_voice) = 0;
617 // Specifies the likelihood that a frame will be declared to contain voice.
618 // A higher value makes it more likely that speech will not be clipped, at
619 // the expense of more noise being detected as voice.
627 virtual int set_likelihood(Likelihood likelihood) = 0;
628 virtual Likelihood likelihood() const = 0;
630 // Sets the |size| of the frames in ms on which the VAD will operate. Larger
631 // frames will improve detection accuracy, but reduce the frequency of
634 // This does not impact the size of frames passed to |ProcessStream()|.
635 virtual int set_frame_size_ms(int size) = 0;
636 virtual int frame_size_ms() const = 0;
639 virtual ~VoiceDetection() {}
641 } // namespace webrtc
643 #endif // WEBRTC_MODULES_AUDIO_PROCESSING_INCLUDE_AUDIO_PROCESSING_H_