* Model the fake voice recorder on a cancellable recording job
FakeVoiceRecorder pushed a whole recording synchronously from a single
suspending call, which made it both weaker and stricter than
DefaultVoiceRecorder.
Every level emission carried the same elapsed time and the emissions were
coalesced by Molecule, so no intermediate recording state was observable and
the level-limit test had to skip half the emissions to compensate. Each level
now advances the test time source and suspends for its share of the recording
duration, so the states arrive one at a time and that magic number is gone.
Call verification moved to the constructor lambdas the rest of the repo uses,
in place of hand-rolled counters and assertCalls. The counters discarded the
argument, so nothing pinned that cancelling a recording stops it with
cancelled = true; that is now asserted. Truth is no longer needed by the
fixture module.
A second startRecord() no longer raises an error where production only warns
and returns, and stopping or deleting terminates an in-flight emission loop
the way cancelling the real recording job does, so a stale recording state can
no longer arrive after the recording has finished.
* Keep each recording's level loop to its own recording
The level loop checked a single isRecording flag, which cannot tell "this
recording was cancelled" from "a later recording has started". After a cancel
followed by a new start, the first loop resumed, saw the flag set again, and
advanced the shared test time source, so the new recording reported the elapsed
time of both. Each recording now carries an id and its loop stops as soon as
that id is no longer the active one.
2 replace all actions have been performed:
- "SPDX-License-Identifier: AGPL-3.0-only" to "SPDX-License-Identifier: AGPL-3.0-only OR LicenseRef-Element-Commercial"
- "Please see LICENSE in the repository root for full details." to "Please see LICENSE files in the repository root for full details."
- New `AudioLevelCalculator` that outputs dB0v rescaled to the [0;1] range.
- `VoiceRecorder` now stores the audio levels sampled while recording, then resamples them to 100 samples to use as waveform preview.
- Waveform data is carried all the way as a `List<Float>` and converted to `List<Int>` in the [0;1024] range as per matrix spec only before sending it.
This is in preparation of further changes to the way the audio level is computed and to allow recording and sending of the waveform. The main reasoning behind the change is twofold:
1) We don't need the precision of Double in our context (we just need a rough indication of the changes in audio level to successfully draw a level meter or a waveform in our UI).
2) Performance: It is true that on 64 bit CPUs single operations involving Floats or Doubles take the same amount of time (i.e one clock cycle). But there are other aspects here that vouch in favor of Floats:
- A float takes half the space in memory compared to a double, so when storing long lists of them this can add up.
- On Android O and greater the ART runtime can "vectorize" certain operations on lists and make use of the CPU's SIMD registers which are generally 128 bits. So by using floats 4 of them can fit and be computed at the same time whilst with doubles only 2 will fit halving the throughput.
References:
- https://source.android.com/docs/core/runtime/improvements
- https://www.slideshare.net/linaroorg/automatic-vectorization-in-art-android-runtime-sfo17216
hayyaksi
Benoit Marty
renovate[bot]
jonnyandrew
David Langley
Marco Romano