From e523ce60360e3da384411d850a18a42a238e4623 Mon Sep 17 00:00:00 2001
From: Travis Ralston <travpc@gmail.com>
Date: Mon, 29 Mar 2021 21:25:06 -0600
Subject: [PATCH] Fix float operations to make a little more sense.

---
 .../voice_messages/LiveRecordingWaveform.tsx  | 12 ++++------
 .../views/voice_messages/Waveform.tsx         |  6 ++---
 src/voice/VoiceRecorder.ts                    | 24 ++++++++-----------
 3 files changed, 18 insertions(+), 24 deletions(-)

diff --git a/src/components/views/voice_messages/LiveRecordingWaveform.tsx b/src/components/views/voice_messages/LiveRecordingWaveform.tsx
index b94eae0f56..e7cab4a5cb 100644
--- a/src/components/views/voice_messages/LiveRecordingWaveform.tsx
+++ b/src/components/views/voice_messages/LiveRecordingWaveform.tsx
@@ -49,16 +49,14 @@ export default class LiveRecordingWaveform extends React.PureComponent<IProps, I
         const bars = arrayFastResample(Array.from(update.waveform), DOWNSAMPLE_TARGET);
         this.setState({
             // 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
-            // point where the average user can still see feedback and be perceived as peaking
-            // when talking "loudly".
-            //
-            // We multiply by 100 because the Waveform component wants values in 0-100 (percentages)
-            heights: bars.map(b => percentageOf(b, 0, 0.50) * 100),
+            // microphone won't send you over 0.6, so we artificially adjust the gain for the
+            // waveform. This results in a slightly more cinematic/animated waveform for the
+            // user.
+            heights: bars.map(b => percentageOf(b, 0, 0.50)),
         });
     };
 
     public render() {
-        return <Waveform heights={this.state.heights} />;
+        return <Waveform relHeights={this.state.heights} />;
     }
 }
diff --git a/src/components/views/voice_messages/Waveform.tsx b/src/components/views/voice_messages/Waveform.tsx
index 4c3edcb927..5fa68dcadc 100644
--- a/src/components/views/voice_messages/Waveform.tsx
+++ b/src/components/views/voice_messages/Waveform.tsx
@@ -18,7 +18,7 @@ import React from "react";
 import {replaceableComponent} from "../../../utils/replaceableComponent";
 
 interface IProps {
-    heights: number[]; // percentages as integers (0-100)
+    relHeights: number[]; // relative heights (0-1)
 }
 
 interface IState {
@@ -37,8 +37,8 @@ export default class Waveform extends React.PureComponent<IProps, IState> {
 
     public render() {
         return <div className='mx_Waveform'>
-            {this.props.heights.map((h, i) => {
-                return <span key={i} style={{height: h + '%'}} className='mx_Waveform_bar' />;
+            {this.props.relHeights.map((h, i) => {
+                return <span key={i} style={{height: (h * 100) + '%'}} className='mx_Waveform_bar' />;
             })}
         </div>;
     }
diff --git a/src/voice/VoiceRecorder.ts b/src/voice/VoiceRecorder.ts
index 6a3d392ce4..50497438ca 100644
--- a/src/voice/VoiceRecorder.ts
+++ b/src/voice/VoiceRecorder.ts
@@ -19,6 +19,7 @@ import encoderPath from 'opus-recorder/dist/encoderWorker.min.js';
 import {MatrixClient} from "matrix-js-sdk/src/client";
 import CallMediaHandler from "../CallMediaHandler";
 import {SimpleObservable} from "matrix-widget-api";
+import {percentageOf} from "../utils/numbers";
 
 const CHANNELS = 1; // stereo isn't important
 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
         // 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.
-        //
-        // We use bytes out of the analyser because floats have weird precision problems
-        // 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);
+        const data = new Float32Array(this.recorderFFT.fftSize);
+        this.recorderFFT.getFloatTimeDomainData(data);
 
-        // Because we're dealing with a uint array we need to do math a bit differently.
-        // If we just `Array.from()` the uint array, we end up with 1s and 0s, which aren't
-        // what we're after. Instead, we have to use a bit of manual looping to correctly end
-        // up with the right values
+        // We can't just `Array.from()` the array because we're dealing with 32bit floats
+        // and the built-in function won't consider that when converting between numbers.
+        // However, the runtime will convert the float32 to a float64 during the math operations
+        // 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[] = [];
         for (let i = 0; i < data.length; i++) {
-            // All we're doing here is inverting the amplitude and putting the metric somewhere
-            // between zero and one. Without the inversion, lower values are "louder", which is
-            // not super helpful.
-            translatedData.push(1 - (data[i] / 128.0));
+            // We're clamping the values so we can do that math operation mentioned above.
+            translatedData.push(percentageOf(data[i], 0, 1));
         }
 
         this.observable.update({