Fix float operations to make a little more sense.

This commit is contained in:
Travis Ralston 2021-03-29 21:25:06 -06:00
parent a848febd3d
commit e523ce6036
3 changed files with 18 additions and 24 deletions

View file

@ -49,16 +49,14 @@ export default class LiveRecordingWaveform extends React.PureComponent<IProps, I
const bars = arrayFastResample(Array.from(update.waveform), DOWNSAMPLE_TARGET); const bars = arrayFastResample(Array.from(update.waveform), DOWNSAMPLE_TARGET);
this.setState({ this.setState({
// The incoming data is between zero and one, but typically even screaming into a // The incoming data is between zero and one, but typically even screaming into a
// microphone won't send you over 0.6, so we "cap" the graph at about 0.50 for a // microphone won't send you over 0.6, so we artificially adjust the gain for the
// point where the average user can still see feedback and be perceived as peaking // waveform. This results in a slightly more cinematic/animated waveform for the
// when talking "loudly". // user.
// heights: bars.map(b => percentageOf(b, 0, 0.50)),
// We multiply by 100 because the Waveform component wants values in 0-100 (percentages)
heights: bars.map(b => percentageOf(b, 0, 0.50) * 100),
}); });
}; };
public render() { public render() {
return <Waveform heights={this.state.heights} />; return <Waveform relHeights={this.state.heights} />;
} }
} }

View file

@ -18,7 +18,7 @@ import React from "react";
import {replaceableComponent} from "../../../utils/replaceableComponent"; import {replaceableComponent} from "../../../utils/replaceableComponent";
interface IProps { interface IProps {
heights: number[]; // percentages as integers (0-100) relHeights: number[]; // relative heights (0-1)
} }
interface IState { interface IState {
@ -37,8 +37,8 @@ export default class Waveform extends React.PureComponent<IProps, IState> {
public render() { public render() {
return <div className='mx_Waveform'> return <div className='mx_Waveform'>
{this.props.heights.map((h, i) => { {this.props.relHeights.map((h, i) => {
return <span key={i} style={{height: h + '%'}} className='mx_Waveform_bar' />; return <span key={i} style={{height: (h * 100) + '%'}} className='mx_Waveform_bar' />;
})} })}
</div>; </div>;
} }

View file

@ -19,6 +19,7 @@ import encoderPath from 'opus-recorder/dist/encoderWorker.min.js';
import {MatrixClient} from "matrix-js-sdk/src/client"; import {MatrixClient} from "matrix-js-sdk/src/client";
import CallMediaHandler from "../CallMediaHandler"; import CallMediaHandler from "../CallMediaHandler";
import {SimpleObservable} from "matrix-widget-api"; import {SimpleObservable} from "matrix-widget-api";
import {percentageOf} from "../utils/numbers";
const CHANNELS = 1; // stereo isn't important const CHANNELS = 1; // stereo isn't important
const SAMPLE_RATE = 48000; // 48khz is what WebRTC uses. 12khz is where we lose quality. const SAMPLE_RATE = 48000; // 48khz is what WebRTC uses. 12khz is where we lose quality.
@ -133,23 +134,18 @@ export class VoiceRecorder {
// The time domain is the input to the FFT, which means we use an array of the same // The time domain is the input to the FFT, which means we use an array of the same
// size. The time domain is also known as the audio waveform. We're ignoring the // size. The time domain is also known as the audio waveform. We're ignoring the
// output of the FFT here (frequency data) because we're not interested in it. // output of the FFT here (frequency data) because we're not interested in it.
// const data = new Float32Array(this.recorderFFT.fftSize);
// We use bytes out of the analyser because floats have weird precision problems this.recorderFFT.getFloatTimeDomainData(data);
// and are slightly more difficult to work with. The bytes are easy to work with,
// which is why we pick them (they're also more precise, but we care less about that).
const data = new Uint8Array(this.recorderFFT.fftSize);
this.recorderFFT.getByteTimeDomainData(data);
// Because we're dealing with a uint array we need to do math a bit differently. // We can't just `Array.from()` the array because we're dealing with 32bit floats
// If we just `Array.from()` the uint array, we end up with 1s and 0s, which aren't // and the built-in function won't consider that when converting between numbers.
// what we're after. Instead, we have to use a bit of manual looping to correctly end // However, the runtime will convert the float32 to a float64 during the math operations
// up with the right values // which is why the loop works below. Note that a `.map()` call also doesn't work
// and will instead return a Float32Array still.
const translatedData: number[] = []; const translatedData: number[] = [];
for (let i = 0; i < data.length; i++) { for (let i = 0; i < data.length; i++) {
// All we're doing here is inverting the amplitude and putting the metric somewhere // We're clamping the values so we can do that math operation mentioned above.
// between zero and one. Without the inversion, lower values are "louder", which is translatedData.push(percentageOf(data[i], 0, 1));
// not super helpful.
translatedData.push(1 - (data[i] / 128.0));
} }
this.observable.update({ this.observable.update({