Feat: Add replay-gain to audio playback

The `RecordingView` struct now includes `gain_db` to represent
replay-gain. This value is read from the recording's metadata (`.json`
sidecar) and passed to the frontend.

The JavaScript in `case_recordings.html` uses this `data-gain-db`
attribute to apply replay-gain using the Web Audio API. This ensures
consistent playback loudness across different recordings. The
implementation uses a shared `AudioContext` to manage resources
efficiently and handles cases where the browser might not support
`AudioContext`.

Additionally, the audio recording on Wear OS now uses
`MediaRecorder.AudioSource.VOICE_RECOGNITION` instead of `MIC`. This
leverages platform noise and echo cancellation, aiming for a cleaner
signal before server-side gain adjustment.
This commit is contained in:
2026-04-27 17:08:10 +02:00
parent 710b60bb16
commit 99f77b666d
3 changed files with 114 additions and 2 deletions
+69
View File
@@ -37,6 +37,13 @@ pub(crate) struct RecordingView {
/// True if the file has been renamed to `<ts>.m4a.failed` by the worker /// True if the file has been renamed to `<ts>.m4a.failed` by the worker
/// after a non-recoverable ffmpeg or whisper error. /// after a non-recoverable ffmpeg or whisper error.
pub(crate) failed: bool, pub(crate) failed: bool,
/// Replay-gain in dB to apply on the client. Pulled from
/// `<stem>.json`'s `loudness.gain_db`. `None` means no gain — pre-
/// loudness recordings, silent files (`-inf` measurement), or
/// recordings whose analyze step failed. The template renders this
/// as `data-gain-db` on the player and the JS multiplies the audio
/// signal accordingly via a Web Audio `GainNode`.
pub(crate) gain_db: Option<f64>,
} }
pub async fn handle_audio( pub async fn handle_audio(
@@ -173,6 +180,7 @@ struct RawRecording {
failed: bool, failed: bool,
transcript: TranscriptState, transcript: TranscriptState,
duration_seconds: Option<u32>, duration_seconds: Option<u32>,
gain_db: Option<f64>,
audio_path: PathBuf, audio_path: PathBuf,
} }
@@ -200,6 +208,10 @@ pub(crate) async fn scan_recordings(case_dir: &Path) -> Vec<RecordingView> {
let stem_path = case_dir.join(stem_file); let stem_path = case_dir.join(stem_file);
let meta = paths::read_recording_meta(&stem_path).await; let meta = paths::read_recording_meta(&stem_path).await;
let duration_seconds = meta.as_ref().and_then(|m| m.duration_seconds); let duration_seconds = meta.as_ref().and_then(|m| m.duration_seconds);
let gain_db = meta
.as_ref()
.and_then(|m| m.loudness.as_ref())
.map(|l| l.gain_db);
let transcript = TranscriptState::from_meta(meta.as_ref()); let transcript = TranscriptState::from_meta(meta.as_ref());
let audio_path = case_dir.join(&filename); let audio_path = case_dir.join(&filename);
@@ -208,6 +220,7 @@ pub(crate) async fn scan_recordings(case_dir: &Path) -> Vec<RecordingView> {
failed, failed,
transcript, transcript,
duration_seconds, duration_seconds,
gain_db,
audio_path, audio_path,
}); });
} }
@@ -249,6 +262,7 @@ pub(crate) async fn scan_recordings(case_dir: &Path) -> Vec<RecordingView> {
duration_seconds: r.duration_seconds, duration_seconds: r.duration_seconds,
transcript: r.transcript, transcript: r.transcript,
failed: r.failed, failed: r.failed,
gain_db: r.gain_db,
} }
}) })
.collect(); .collect();
@@ -330,6 +344,61 @@ mod tests {
assert!(parse_range("bytes=-0", 1000).is_none()); assert!(parse_range("bytes=-0", 1000).is_none());
} }
/// Recording with a `loudness` block in its sidecar surfaces
/// `gain_db` on the `RecordingView`. Seeding a full
/// `RecordingMeta` directly via serde_json instead of going
/// through `write_recording_meta_sync`, since that helper does
/// not (yet) carry a loudness parameter.
#[tokio::test]
async fn scan_reads_gain_db_from_loudness() {
let dir = tempdir().unwrap();
let stem = "2026-04-19T13-00-00Z";
tokio::fs::write(dir.path().join(format!("{stem}.m4a")), b"audio")
.await
.unwrap();
let meta = doctate_common::RecordingMeta {
transcript: doctate_common::Transcript::Content { text: "hi".into() },
duration_seconds: Some(30),
loudness: Some(doctate_common::Loudness {
mean_db: -22.0,
max_db: -15.0,
gain_db: 6.0,
}),
};
let bytes = serde_json::to_vec(&meta).unwrap();
tokio::fs::write(dir.path().join(format!("{stem}.json")), bytes)
.await
.unwrap();
let got = scan_recordings(dir.path()).await;
assert_eq!(got.len(), 1);
assert_eq!(got[0].gain_db, Some(6.0));
assert_eq!(got[0].duration_seconds, Some(30));
}
/// Pre-loudness sidecars (worker wrote meta before this feature
/// existed) carry no `loudness` field, deserialize as `None`
/// (thanks to `#[serde(default)]`), and surface `gain_db: None`
/// on the view. The frontend then renders `data-gain-db="0"` and
/// the player skips the Web-Audio routing.
#[tokio::test]
async fn scan_yields_none_gain_db_for_meta_without_loudness() {
let dir = tempdir().unwrap();
tokio::fs::write(dir.path().join("2026-04-19T14-00-00Z.m4a"), b"audio")
.await
.unwrap();
crate::paths::write_recording_meta_sync(
dir.path(),
"2026-04-19T14-00-00Z",
doctate_common::Transcript::Silent,
Some(7),
);
let got = scan_recordings(dir.path()).await;
assert_eq!(got.len(), 1);
assert_eq!(got[0].gain_db, None);
}
#[tokio::test] #[tokio::test]
async fn scan_handles_malformed_meta_sidecar() { async fn scan_handles_malformed_meta_sidecar() {
let dir = tempdir().unwrap(); let dir = tempdir().unwrap();
+40 -1
View File
@@ -143,7 +143,7 @@ try {
<div class="recording{% if rec.failed %} failed-row{% endif %}"> <div class="recording{% if rec.failed %} failed-row{% endif %}">
<div class="recording-head"> <div class="recording-head">
<time class="rec-time" datetime="{{ rec.recorded_at_iso }}">{{ rec.recorded_at_iso }}</time> <time class="rec-time" datetime="{{ rec.recorded_at_iso }}">{{ rec.recorded_at_iso }}</time>
<div class="player" data-src="/web/audio/{{ slug }}/{{ case_id }}/{{ rec.filename }}"{% match rec.duration_seconds %}{% when Some with (d) %} data-duration="{{ d }}"{% when None %}{% endmatch %}> <div class="player" data-src="/web/audio/{{ slug }}/{{ case_id }}/{{ rec.filename }}"{% match rec.duration_seconds %}{% when Some with (d) %} data-duration="{{ d }}"{% when None %}{% endmatch %}{% match rec.gain_db %}{% when Some with (g) %} data-gain-db="{{ g }}"{% when None %}{% endmatch %}>
<button type="button" class="play" aria-label="Abspielen"></button> <button type="button" class="play" aria-label="Abspielen"></button>
<input type="range" class="seek" min="0" max="1000" value="0" step="1" aria-label="Position"> <input type="range" class="seek" min="0" max="1000" value="0" step="1" aria-label="Position">
<span class="time">0:00 / 0:00</span> <span class="time">0:00 / 0:00</span>
@@ -247,6 +247,25 @@ try {
return `${m}:${sec.toString().padStart(2, '0')}`; return `${m}:${sec.toString().padStart(2, '0')}`;
}; };
// Shared AudioContext for replay-gain. Lazy-created on the first
// player that actually needs gain (Safari requires user-gesture
// creation); shared across all players to avoid hitting the
// per-page context limit. `null` if the browser does not expose
// an AudioContext at all — falls back to the plain <audio> path.
let sharedCtx = null;
const ensureCtx = () => {
if (!sharedCtx) {
const Ctx = window.AudioContext || window.webkitAudioContext;
if (Ctx) sharedCtx = new Ctx();
}
if (sharedCtx && sharedCtx.state === 'suspended') {
// Some Safari builds keep the context suspended even when
// created from a user gesture. Resume is a no-op when running.
sharedCtx.resume();
}
return sharedCtx;
};
document.querySelectorAll('.player').forEach(p => { document.querySelectorAll('.player').forEach(p => {
const src = p.dataset.src; const src = p.dataset.src;
const btn = p.querySelector('.play'); const btn = p.querySelector('.play');
@@ -287,6 +306,26 @@ try {
if (audio) return audio; if (audio) return audio;
audio = new Audio(src); audio = new Audio(src);
audio.preload = 'metadata'; audio.preload = 'metadata';
// Replay-gain: route through Web Audio with a GainNode so the
// recording plays at the consistent target loudness regardless
// of which client recorded it. `createMediaElementSource`
// *redirects* the audio element's output through the graph —
// forgetting the `connect(ctx.destination)` would silence the
// player. Skipped entirely when gain is 0, missing, or
// negative (older meta sidecars, silent files, analyze errors)
// so the default <audio> path still works.
const gainDb = parseFloat(p.dataset.gainDb);
if (isFinite(gainDb) && gainDb > 0) {
const ctx = ensureCtx();
if (ctx) {
const mediaSrc = ctx.createMediaElementSource(audio);
const gainNode = ctx.createGain();
gainNode.gain.value = Math.pow(10, gainDb / 20);
mediaSrc.connect(gainNode).connect(ctx.destination);
}
}
audio.addEventListener('loadedmetadata', () => { audio.addEventListener('loadedmetadata', () => {
if (isFinite(audio.duration) && audio.duration > 0) duration = audio.duration; if (isFinite(audio.duration) && audio.duration > 0) duration = audio.duration;
tryApplyPendingSeek(); tryApplyPendingSeek();
@@ -26,7 +26,11 @@ class AudioRecorder(private val context: Context) {
require(recorder == null) { "recorder already active" } require(recorder == null) { "recorder already active" }
outFile.parentFile?.takeIf { !it.exists() }?.mkdirs() outFile.parentFile?.takeIf { !it.exists() }?.mkdirs()
val r = newRecorderInstance().apply { val r = newRecorderInstance().apply {
setAudioSource(MediaRecorder.AudioSource.MIC) // VOICE_RECOGNITION lowers raw level vs MIC but applies the
// platform's noise/echo cleaning. Combined with the
// server-side replay-gain pass the playback ends up at
// target loudness with cleaner SNR than raw MIC + post-gain.
setAudioSource(MediaRecorder.AudioSource.VOICE_RECOGNITION)
setOutputFormat(MediaRecorder.OutputFormat.MPEG_4) setOutputFormat(MediaRecorder.OutputFormat.MPEG_4)
setAudioEncoder(MediaRecorder.AudioEncoder.AAC) setAudioEncoder(MediaRecorder.AudioEncoder.AAC)
setAudioSamplingRate(16_000) setAudioSamplingRate(16_000)