Skip to content

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.

fix(upload): survive poisoned TLS connections; cap upload diarization speaker clusters (#1496)

Joshua6d2d30340a06
diff --git a/src/components/notes/UploadAudioView.tsx b/src/components/notes/UploadAudioView.tsx
index 733aba9c7..4ed04dc43 100644
--- a/src/components/notes/UploadAudioView.tsx
+++ b/src/components/notes/UploadAudioView.tsx
@@ -29,6 +29,7 @@ import {
findDefaultFolder,
findVideosFolder,
DOWNLOAD_ERROR_KEYS,
+ transcriptionErrorKey,
MEETINGS_FOLDER_NAME,
} from "./shared";
import { useAuth } from "../../hooks/useAuth";
@@ -135,7 +136,9 @@ export default function UploadAudioView({ onNoteCreated, onOpenSettings }: Uploa
durationSeconds?: number | null;
} | null>(null);
const [result, setResult] = useState<string | null>(null);
- const [partialWarning, setPartialWarning] = useState(false);
+ const [partialWarning, setPartialWarning] = useState<{ failed: number; total: number } | null>(
+ null
+ );
const [noteId, setNoteId] = useState<number | null>(null);
const [error, setError] = useState<string | null>(null);
const [isDragOver, setIsDragOver] = useState(false);
@@ -584,7 +587,7 @@ export default function UploadAudioView({ onNoteCreated, onOpenSettings }: Uploa
setState("idle");
setFile(null);
setResult(null);
- setPartialWarning(false);
+ setPartialWarning(null);
setNoteId(null);
setError(null);
setProgress(0);
@@ -671,7 +674,11 @@ export default function UploadAudioView({ onNoteCreated, onOpenSettings }: Uploa
if (res.success && res.text) {
setProgress(100);
setResult(res.text);
- setPartialWarning(!!res.warning);
+ setPartialWarning(
+ res.failedChunks && res.totalChunks
+ ? { failed: res.failedChunks, total: res.totalChunks }
+ : null
+ );
let title: string;
if (currentFile.fromUrl) {
@@ -705,9 +712,10 @@ export default function UploadAudioView({ onNoteCreated, onOpenSettings }: Uploa
setState("complete");
} else {
setProgress(0);
+ const errorKey = transcriptionErrorKey(res);
setError(
- res.code === "NO_SPEECH_DETECTED"
- ? t("notes.upload.noSpeechDetected")
+ errorKey
+ ? t(`notes.upload.${errorKey}`)
: res.error || t("notes.upload.transcriptionFailed")
);
setState("error");
@@ -718,7 +726,12 @@ export default function UploadAudioView({ onNoteCreated, onOpenSettings }: Uploa
if (progressCleanupRef.current) progressCleanupRef.current();
progressCleanupRef.current = null;
setProgress(0);
- setError(err instanceof Error ? err.message : t("notes.upload.errorOccurred"));
+ const errorKey = transcriptionErrorKey(err);
+ if (errorKey) {
+ setError(t(`notes.upload.${errorKey}`));
+ } else {
+ setError(err instanceof Error ? err.message : t("notes.upload.errorOccurred"));
+ }
setState("error");
}
};
@@ -1289,7 +1302,14 @@ export default function UploadAudioView({ onNoteCreated, onOpenSettings }: Uploa
{diarizationEnabled && diarizationModelsReady && (
<div className="mt-2">
+ <label
+ htmlFor="upload-num-speakers"
+ className="block text-xs font-medium text-foreground/50"
+ >
+ {t("notes.upload.numSpeakersLabel")}
+ </label>
<input
+ id="upload-num-speakers"
type="number"
min="2"
max={MAX_SPEAKER_COUNT}
@@ -1304,12 +1324,14 @@ export default function UploadAudioView({ onNoteCreated, onOpenSettings }: Uploa
setDiarizationNumSpeakers(String(isNaN(n) ? "" : n));
}}
placeholder={t("notes.upload.numSpeakersPlaceholder")}
- aria-label={t("notes.upload.numSpeakersPlaceholder")}
className={cn(
uploadFieldClass,
- "w-full h-8 px-2.5 [appearance:textfield] [&::-webkit-outer-spin-button]:appearance-none [&::-webkit-inner-spin-button]:appearance-none"
+ "mt-1 w-full h-8 px-2.5 [appearance:textfield] [&::-webkit-outer-spin-button]:appearance-none [&::-webkit-inner-spin-button]:appearance-none"
)}
/>
+ <p className="text-[10px] text-foreground/25 mt-1.5">
+ {t("notes.upload.numSpeakersHint")}
+ </p>
</div>
)}
</div>
@@ -1807,9 +1829,9 @@ function FolderSelect({
}
interface CompleteViewProps {
- t: (key: string) => string;
+ t: (key: string, options?: Record<string, unknown>) => string;
result: string;
- partialWarning: boolean;
+ partialWarning: { failed: number; total: number } | null;
folders: FolderItem[];
selectedFolderId: string;
handleFolderChange: (val: string) => void;
@@ -1878,7 +1900,10 @@ function CompleteView({
{partialWarning && (
<p className="text-xs text-destructive/50 max-w-[240px] text-center mb-4 -mt-2">
- {t("notes.upload.partialWarning")}
+ {t("notes.upload.partialWarningCount", {
+ failed: partialWarning.failed,
+ total: partialWarning.total,
+ })}
</p>
)}
diff --git a/src/components/notes/overview/OverviewNoteList.tsx b/src/components/notes/overview/OverviewNoteList.tsx
index cba8d4155..4cc8f313b 100644
--- a/src/components/notes/overview/OverviewNoteList.tsx
+++ b/src/components/notes/overview/OverviewNoteList.tsx
@@ -77,7 +77,10 @@ export function OverviewNoteList({
onClick={() => onOpenNote(note.id)}
className="w-full flex items-center gap-3 px-2 py-2 -mx-2 rounded-md text-left hover:bg-foreground/4 dark:hover:bg-white/4 transition-colors duration-150 focus:outline-none focus-visible:ring-1 focus-visible:ring-ring/30"
>
- <FileText size={14} className="text-foreground/30 dark:text-foreground/20 shrink-0" />
+ <FileText
+ size={14}
+ className="text-foreground/30 dark:text-foreground/20 shrink-0"
+ />
<span className="text-[13px] text-foreground/85 truncate flex-1">
{note.title || t("notes.list.untitled")}
</span>
diff --git a/src/components/notes/shared.ts b/src/components/notes/shared.ts
index de0760dd6..edfb0cb67 100644
--- a/src/components/notes/shared.ts
+++ b/src/components/notes/shared.ts
@@ -27,6 +27,21 @@ export const DOWNLOAD_ERROR_KEYS: Record<string, string> = {
SSRF_BLOCKED: "urlDownloadFailed",
};
+// Transcription error codes → notes.upload.* i18n keys. Codes absent here fall
+// back to the raw main-process message.
+const TRANSCRIPTION_ERROR_KEYS: Record<string, string> = {
+ NO_SPEECH_DETECTED: "noSpeechDetected",
+ CHUNK_LOSS_EXCEEDED: "chunkLossExceeded",
+};
+
+// A coded failure arrives either as a returned result (BYOK, local) or as a
+// thrown error — OpenWhispr Cloud rethrows it through withSessionRefresh — so
+// every call site resolves the key from whichever shape it is holding.
+export function transcriptionErrorKey(failure: unknown): string | undefined {
+ const code = (failure as { code?: string } | null | undefined)?.code;
+ return code ? TRANSCRIPTION_ERROR_KEYS[code] : undefined;
+}
+
export const notesInputClass = cn(
"w-full h-8 px-3 rounded-md text-xs",
"bg-foreground/3 dark:bg-white/4 border border-border/30 dark:border-white/6",
diff --git a/src/helpers/audioManager.js b/src/helpers/audioManager.js
index 543ff2496..47adc3e32 100644
--- a/src/helpers/audioManager.js
+++ b/src/helpers/audioManager.js
@@ -2715,6 +2715,11 @@ registerProcessor("pcm-streaming-processor", PCMStreamingProcessor);
if (!res.success) {
const err = new Error(res.error || "Cloud transcription failed");
err.code = res.code;
+ // The recording is kept by saveFailedTranscription, so point the user
+ // at History rather than leaving them with a raw main-process string.
+ if (res.code === "CHUNK_LOSS_EXCEEDED") {
+ err.messageKey = "hooks.audioRecording.errorDescriptions.chunkLoss";
+ }
throw err;
}
return res;
diff --git a/src/helpers/cloudChunkPolicy.js b/src/helpers/cloudChunkPolicy.js
index 47156736f..34e8e177f 100644
--- a/src/helpers/cloudChunkPolicy.js
+++ b/src/helpers/cloudChunkPolicy.js
@@ -1,4 +1,6 @@
const { createAbortError } = require("./abortError");
+const { formatTimestamp } = require("./speakerMerge");
+const { i18nMain } = require("./i18nMain");
// Retry/backoff/concurrency policy for the chunked cloud upload path (#1326).
// Kept free of electron imports so the rules stay unit-testable.
@@ -29,6 +31,21 @@ const CLOUD_CHUNK_BACKOFF_JITTER_MS = 1_000;
// sacrifice its peers — and its peers' retries in turn.
const CLOUD_POOL_TEARDOWN_COOLDOWN_MS = 30_000;
+// A teardown aborts every sibling still on the wire; those failures are
+// self-inflicted, so the sibling's attempt is refunded rather than charged.
+// Bounded per chunk so repeated teardowns can't retry forever.
+const CLOUD_CHUNK_MAX_TEARDOWN_REFUNDS = 3;
+
+// Above this failed/total ratio the "transcript" is more hole than content;
+// returning success would silently persist a fragment (15/19 chunks lost once
+// shipped a 16-minute transcript of a 73-minute file as a normal note).
+const CLOUD_CHUNK_MAX_LOSS_RATIO = 0.5;
+
+// A 422 NO_SPEECH response is a valid empty segment, not an upload failure.
+// Keep it distinct from null (failed) and response objects (transcribed) so
+// silence neither consumes the loss budget nor creates a missing-audio marker.
+const SILENT_CHUNK = Symbol("silent-cloud-chunk");
+
// Codes that doom the whole job, not just one chunk: retrying the siblings can
// only burn time and quota.
const FATAL_CHUNK_CODES = new Set(["AUTH_EXPIRED", "LIMIT_REACHED"]);
@@ -48,16 +65,84 @@ function isNetworkLevelFailure(err, { timedOut = false } = {}) {
return timedOut || (!err.statusCode && !err.code);
}
+// Fatal TLS alerts and HTTP/2/QUIC protocol errors poison every stream on the
+// connection, so these force the teardown through the cooldown — which exists
+// to absorb bursts of ordinary failures, not to keep a provably sick pool
+// alive. Chromium's fetch rejections carry the net::ERR_* name only in the
+// message, never a structured code.
+const POISONED_CONNECTION_RE = /net::ERR_(SSL|HTTP2|QUIC)_/;
+
+function isConnectionPoisoningFailure(err) {
+ return POISONED_CONNECTION_RE.test(err?.message ?? "");
+}
+
+// True when the failure was (very likely) our own pool teardown aborting this
+// chunk's in-flight body: a teardown happened during the attempt and the
+// request died at the transport level. An HTTP answer or business code proves
+// the request outlived the teardown, and a timeout is the chunk's own failure.
+function isTeardownCollateral(err, { timedOut = false, teardownsDuringAttempt = 0 } = {}) {
+ return teardownsDuringAttempt > 0 && !timedOut && isNetworkLevelFailure(err, { timedOut });
+}
+
function createTeardownGate(cooldownMs = CLOUD_POOL_TEARDOWN_COOLDOWN_MS, now = Date.now) {
let lastAt = -Infinity;
- return () => {
+ return (force = false) => {
const at = now();
- if (at - lastAt < cooldownMs) return false;
+ if (!force && at - lastAt < cooldownMs) return false;
lastAt = at;
return true;
};
}
+function summarizeChunkResults(results) {
+ const responses = [];
+ let failedChunks = 0;
+ let silentChunks = 0;
+
+ for (const result of results) {
+ if (result == null) {
+ failedChunks++;
+ } else if (result === SILENT_CHUNK) {
+ silentChunks++;
+ } else {
+ responses.push(result);
+ }
+ }
+
+ return { responses, failedChunks, silentChunks };
+}
+
+// Joins surviving chunk texts in order and marks every failed span with an
+// explicit gap (consecutive failures collapse into one marker), so a partial
+// transcript reads as partial instead of seamlessly wrong. Silent segments are
+// omitted, and a known input duration clamps the final gap to its real end. The
+// marker is read by users and pasted by dictation, so it is localized like the
+// [Speaker N] labels in transcriptFormatter.
+function assembleChunkTranscript(results, segmentDurationSeconds, totalDurationSeconds) {
+ const pieces = [];
+ for (let i = 0; i < results.length; i++) {
+ if (results[i] === SILENT_CHUNK) continue;
+ if (results[i] != null) {
+ pieces.push(results[i].text);
+ continue;
+ }
+ const gapStart = i;
+ while (i + 1 < results.length && results[i + 1] == null) i++;
+ const gapEnd = (i + 1) * segmentDurationSeconds;
+ pieces.push(
+ i18nMain.t("transcript.missingAudio", {
+ from: formatTimestamp(gapStart * segmentDurationSeconds),
+ to: formatTimestamp(
+ Number.isFinite(totalDurationSeconds) && totalDurationSeconds > 0
+ ? Math.min(gapEnd, totalDurationSeconds)
+ : gapEnd
+ ),
+ })
+ );
+ }
+ return pieces.join(" ").replace(/\s+/g, " ").trim();
+}
+
function chunkRetryDelayMs(attempt, random = Math.random) {
const base = Math.min(
CLOUD_CHUNK_BACKOFF_BASE_MS * CLOUD_CHUNK_BACKOFF_FACTOR ** (attempt - 1),
@@ -146,9 +231,16 @@ module.exports = {
CLOUD_CHUNK_MAX_ATTEMPTS,
CLOUD_CHUNK_GLOBAL_CONCURRENCY,
CLOUD_POOL_TEARDOWN_COOLDOWN_MS,
+ CLOUD_CHUNK_MAX_TEARDOWN_REFUNDS,
+ CLOUD_CHUNK_MAX_LOSS_RATIO,
+ SILENT_CHUNK,
FATAL_CHUNK_CODES,
isTransientChunkError,
isNetworkLevelFailure,
+ isConnectionPoisoningFailure,
+ isTeardownCollateral,
+ summarizeChunkResults,
+ assembleChunkTranscript,
chunkRetryDelayMs,
abortableSleep,
createTeardownGate,
diff --git a/src/helpers/diarizationPolicy.js b/src/helpers/diarizationPolicy.js
new file mode 100644
index 000000000..39c47c2c1
--- /dev/null
+++ b/src/helpers/diarizationPolicy.js
@@ -0,0 +1,63 @@
+// Clustering policy for the upload/batch diarization path. Kept free of
+// electron imports so the rules stay unit-testable (pattern: cloudChunkPolicy).
+
+// sherpa-onnx's cluster threshold is a cosine distance: higher merges more.
+// 0.55 suits short clean audio, but over a long single-mic recording one
+// speaker's embeddings drift wider than that and agglomerative clustering
+// stops merging early — a 73-minute voice memo produced 46 "speakers". Ramp
+// the threshold with duration instead of pretending one constant fits both.
+const DEFAULT_CLUSTER_THRESHOLD = 0.55;
+const LONG_AUDIO_CLUSTER_THRESHOLD = 0.8;
+const THRESHOLD_RAMP_START_SECONDS = 15 * 60;
+const THRESHOLD_RAMP_END_SECONDS = 60 * 60;
+
+// A cluster with under a second of total speech is embedding noise, not a
+// person — and the character-proportional merge hands every surviving cluster
+// at least one sentence, so each blip becomes a phantom [Speaker N].
+const MIN_CLUSTER_TOTAL_SECONDS = 1;
+
+function clusterThresholdForDuration(durationSeconds) {
+ if (!Number.isFinite(durationSeconds) || durationSeconds <= THRESHOLD_RAMP_START_SECONDS) {
+ return DEFAULT_CLUSTER_THRESHOLD;
+ }
+ if (durationSeconds >= THRESHOLD_RAMP_END_SECONDS) return LONG_AUDIO_CLUSTER_THRESHOLD;
+ const progress =
+ (durationSeconds - THRESHOLD_RAMP_START_SECONDS) /
+ (THRESHOLD_RAMP_END_SECONDS - THRESHOLD_RAMP_START_SECONDS);
+ return (
+ DEFAULT_CLUSTER_THRESHOLD +
+ (LONG_AUDIO_CLUSTER_THRESHOLD - DEFAULT_CLUSTER_THRESHOLD) * progress
+ );
+}
+
+function resolveClusterThreshold(durationSeconds, requestedThreshold) {
+ const automaticThreshold = clusterThresholdForDuration(durationSeconds);
+ if (requestedThreshold == null || requestedThreshold === "") return automaticThreshold;
+ const parsedThreshold = Number(requestedThreshold);
+ if (!Number.isFinite(parsedThreshold)) return automaticThreshold;
+ return Math.min(1, Math.max(0, parsedThreshold));
+}
+
+function dropNegligibleClusters(segments, minTotalSeconds = MIN_CLUSTER_TOTAL_SECONDS) {
+ if (!segments?.length) return segments;
+ const totals = new Map();
+ for (const s of segments) {
+ totals.set(s.speaker, (totals.get(s.speaker) || 0) + (s.end - s.start));
+ }
+ const keep = new Set(
+ [...totals].filter(([, total]) => total >= minTotalSeconds).map(([speaker]) => speaker)
+ );
+ if (keep.size === 0 || keep.size === totals.size) return segments;
+ return segments.filter((s) => keep.has(s.speaker));
+}
+
+module.exports = {
+ DEFAULT_CLUSTER_THRESHOLD,
+ LONG_AUDIO_CLUSTER_THRESHOLD,
+ THRESHOLD_RAMP_START_SECONDS,
+ THRESHOLD_RAMP_END_SECONDS,
+ MIN_CLUSTER_TOTAL_SECONDS,
+ clusterThresholdForDuration,
+ resolveClusterThreshold,
+ dropNegligibleClusters,
+};
diff --git a/src/helpers/ffmpegUtils.js b/src/helpers/ffmpegUtils.js
index d051013be..643069b10 100644
--- a/src/helpers/ffmpegUtils.js
+++ b/src/helpers/ffmpegUtils.js
@@ -266,6 +266,12 @@ function computeFloat32RMS(float32Buffer) {
return Math.sqrt(sumSquares / numSamples);
}
+function parseFfmpegDuration(stderr) {
+ const match = stderr?.match(/Duration:\s*(\d+):([0-5]\d):([0-5]\d(?:\.\d+)?)/);
+ if (!match) return null;
+ return Number(match[1]) * 3600 + Number(match[2]) * 60 + Number(match[3]);
+}
+
function splitAudioFile(inputPath, outputDir, options = {}) {
const { segmentDuration = 600, audioBitrate = "128k", signal } = options;
@@ -363,8 +369,9 @@ function splitAudioFile(inputPath, outputDir, options = {}) {
return;
}
- debugLogger.debug("FFmpeg split complete", { chunkCount: chunks.length });
- resolve(chunks);
+ const durationSeconds = parseFfmpegDuration(stderr);
+ debugLogger.debug("FFmpeg split complete", { chunkCount: chunks.length, durationSeconds });
+ resolve({ chunkPaths: chunks, durationSeconds });
});
});
}
@@ -452,6 +459,7 @@ module.exports = {
parseWavFormat,
convertToWav,
splitAudioFile,
+ parseFfmpegDuration,
wavToFloat32Samples,
computeFloat32RMS,
mergeAudioSegments,
diff --git a/src/helpers/ipcHandlers.js b/src/helpers/ipcHandlers.js
index 4ba741bd1..d1305d53d 100644
--- a/src/helpers/ipcHandlers.js
+++ b/src/helpers/ipcHandlers.js
@@ -133,36 +133,56 @@ const {
CLOUD_UPLOAD_TIMEOUT_MS,
CLOUD_CHUNK_MAX_ATTEMPTS,
CLOUD_CHUNK_GLOBAL_CONCURRENCY,
+ CLOUD_CHUNK_MAX_TEARDOWN_REFUNDS,
+ CLOUD_CHUNK_MAX_LOSS_RATIO,
+ SILENT_CHUNK,
FATAL_CHUNK_CODES,
isTransientChunkError,
isNetworkLevelFailure,
+ isConnectionPoisoningFailure,
+ isTeardownCollateral,
+ summarizeChunkResults,
+ assembleChunkTranscript,
chunkRetryDelayMs,
abortableSleep,
createTeardownGate,
createUploadSlots,
} = require("./cloudChunkPolicy");
-// Every cloud upload rides a dedicated in-memory session so stalled large
-// bodies can't wedge the HTTP/2 connection the rest of the app multiplexes
-// over, and a failed attempt can drop the pool without collateral damage
-// outside the upload path (#1326).
-const CLOUD_UPLOAD_SESSION_PARTITION = "ow-cloud-uploads";
+// Chunk retries need their own connection pool: recovering a wedged chunk pool
+// must not abort an unrelated inline upload that has no collateral retry path.
+const CLOUD_CHUNK_UPLOAD_SESSION_PARTITION = "ow-cloud-chunk-uploads";
+const CLOUD_INLINE_UPLOAD_SESSION_PARTITION = "ow-cloud-uploads";
const cloudUploadSlots = createUploadSlots(CLOUD_CHUNK_GLOBAL_CONCURRENCY);
const shouldDropUploadPool = createTeardownGate();
-let cloudUploadSession = null;
+const cloudUploadSessions = new Map();
-function getCloudUploadSession() {
- if (!cloudUploadSession) {
- cloudUploadSession = session.fromPartition(CLOUD_UPLOAD_SESSION_PARTITION);
- applyOpenWhisprOriginHeader(cloudUploadSession);
+function getCloudUploadSession(partition) {
+ if (!cloudUploadSessions.has(partition)) {
+ const uploadSession = session.fromPartition(partition);
+ applyOpenWhisprOriginHeader(uploadSession);
+ cloudUploadSessions.set(partition, uploadSession);
}
- return cloudUploadSession;
+ return cloudUploadSessions.get(partition);
}
-async function dropUploadConnections() {
- if (!shouldDropUploadPool()) return;
+function getChunkCloudUploadSession() {
+ return getCloudUploadSession(CLOUD_CHUNK_UPLOAD_SESSION_PARTITION);
+}
+
+function getInlineCloudUploadSession() {
+ return getCloudUploadSession(CLOUD_INLINE_UPLOAD_SESSION_PARTITION);
+}
+
+// Counts initiated pool drops so a chunk can tell whether its failure was
+// collateral from a teardown that happened while its body was on the wire.
+let uploadPoolTeardowns = 0;
+
+async function dropUploadConnections(force = false) {
+ if (!shouldDropUploadPool(force)) return;
+ uploadPoolTeardowns++;
try {
- await getCloudUploadSession().closeAllConnections();
+ await getChunkCloudUploadSession().closeAllConnections();
} catch {
// pool teardown is best-effort
}
@@ -333,7 +353,7 @@ async function chunkedCloudTranscribe({
try {
onProgress?.({ stage: "splitting", chunksTotal: 0, chunksCompleted: 0 });
- const chunkPaths = await splitAudioFile(inputPath, chunkDir, {
+ const { chunkPaths, durationSeconds } = await splitAudioFile(inputPath, chunkDir, {
segmentDuration,
signal: jobSignal,
});
@@ -343,51 +363,83 @@ async function chunkedCloudTranscribe({
const url = new URL(`${apiUrl}/api/transcribe`);
const results = new Array(totalChunks).fill(null);
- const failureCodes = new Set();
let fatalError = null;
let completedCount = 0;
const transcribeChunk = async (index) => {
- for (let attempt = 1; ; attempt++) {
+ let attempt = 1;
+ let teardownRefunds = CLOUD_CHUNK_MAX_TEARDOWN_REFUNDS;
+ while (true) {
if (jobSignal.aborted) throw createAbortError();
// Held only while a body is on the wire, so the backoff below never
// occupies a slot and queue time never eats the upload timeout.
const releaseSlot = await cloudUploadSlots.acquire(jobSignal);
const timeoutSignal = AbortSignal.timeout(CLOUD_UPLOAD_TIMEOUT_MS);
+ const teardownsAtStart = uploadPoolTeardowns;
let failure = null;
+ let timedOut = false;
+ let collateral = false;
try {
- const { body, boundary } = buildMultipartBody(
- fs.readFileSync(chunkPaths[index]),
- path.basename(chunkPaths[index]),
- "audio/mpeg",
- multipartFields
- );
- const data = await postMultipart(url, body, boundary, policyHeaders, {
- signal: AbortSignal.any([jobSignal, timeoutSignal]),
- session: getCloudUploadSession(),
- });
- results[index] = interpretTranscribeResponse(data);
- } catch (err) {
- failure = err;
+ try {
+ const { body, boundary } = buildMultipartBody(
+ fs.readFileSync(chunkPaths[index]),
+ path.basename(chunkPaths[index]),
+ "audio/mpeg",
+ multipartFields
+ );
+ const data = await postMultipart(url, body, boundary, policyHeaders, {
+ signal: AbortSignal.any([jobSignal, timeoutSignal]),
+ session: getChunkCloudUploadSession(),
+ });
+ results[index] = interpretTranscribeResponse(data);
+ } catch (err) {
+ failure = err;
+ timedOut = timeoutSignal.aborted;
+ collateral = isTeardownCollateral(err, {
+ timedOut,
+ teardownsDuringAttempt: uploadPoolTeardowns - teardownsAtStart,
+ });
+ // Drop the pool while still holding the slot — released first, a
+ // queued sibling is admitted onto the pool microseconds before
+ // closeAllConnections() kills it and burns an attempt it never
+ // owned. A fatal TLS/protocol alert proves the pool is poisoned,
+ // so that drop is forced through the cooldown gate.
+ if (!jobSignal.aborted && !collateral) {
+ const poisoned = isConnectionPoisoningFailure(err);
+ if (poisoned || isNetworkLevelFailure(err, { timedOut })) {
+ await dropUploadConnections(poisoned);
+ }
+ }
+ }
} finally {
releaseSlot();
}
if (!failure) break;
+ if (failure.code === "NO_SPEECH_DETECTED") {
+ results[index] = SILENT_CHUNK;
+ break;
+ }
+
if (jobSignal.aborted) throw createAbortError();
- const timedOut = timeoutSignal.aborted;
+ if (collateral && teardownRefunds > 0) {
+ teardownRefunds--;
+ debugLogger.warn(`Chunk ${index} attempt ${attempt} killed by pool teardown, refunded`, {
+ error: failure.message,
+ });
+ await abortableSleep(chunkRetryDelayMs(1), jobSignal);
+ continue;
+ }
if (attempt >= CLOUD_CHUNK_MAX_ATTEMPTS || !(timedOut || isTransientChunkError(failure))) {
throw failure;
}
- // No HTTP answer ever arrived — treat the pool as wedged and drop it so
- // the retry dials a fresh connection instead of re-entering the dying one.
- if (isNetworkLevelFailure(failure, { timedOut })) await dropUploadConnections();
debugLogger.warn(`Chunk ${index} attempt ${attempt} failed, retrying`, {
error: failure.message,
timedOut,
});
await abortableSleep(chunkRetryDelayMs(attempt), jobSignal);
+ attempt++;
}
completedCount++;
@@ -409,7 +461,6 @@ async function chunkedCloudTranscribe({
abortJob();
return;
}
- if (err.code) failureCodes.add(err.code);
debugLogger.warn(`Chunk ${index} failed`, { error: err.message, code: err.code });
})
)
@@ -420,9 +471,9 @@ async function chunkedCloudTranscribe({
}
if (fatalError) throw fatalError;
- const succeeded = results.filter((r) => r !== null);
- if (succeeded.length === 0) {
- if (failureCodes.size === 1 && failureCodes.has("NO_SPEECH_DETECTED")) {
+ const { responses, failedChunks: failed, silentChunks } = summarizeChunkResults(results);
+ if (responses.length === 0) {
+ if (silentChunks === totalChunks) {
throw Object.assign(new Error("No speech detected in audio"), {
code: "NO_SPEECH_DETECTED",
});
@@ -430,19 +481,24 @@ async function chunkedCloudTranscribe({
throw new Error("All chunks failed to transcribe");
}
- const text = results
- .filter((r) => r !== null)
- .map((r) => r.text)
- .join(" ")
- .replace(/\s+/g, " ")
- .trim();
+ if (failed / totalChunks > CLOUD_CHUNK_MAX_LOSS_RATIO) {
+ throw Object.assign(new Error(`${failed} of ${totalChunks} audio segments were lost`), {
+ code: "CHUNK_LOSS_EXCEEDED",
+ });
+ }
- const failed = totalChunks - succeeded.length;
+ const text = assembleChunkTranscript(results, segmentDuration, durationSeconds);
return {
text,
- responses: succeeded,
- lastResponse: succeeded[succeeded.length - 1],
- ...(failed > 0 ? { warning: `${failed} of ${totalChunks} chunks failed` } : {}),
+ responses,
+ lastResponse: responses[responses.length - 1],
+ ...(failed > 0
+ ? {
+ warning: `${failed} of ${totalChunks} chunks failed`,
+ failedChunks: failed,
+ totalChunks,
+ }
+ : {}),
};
} finally {
signal?.removeEventListener("abort", abortJob);
@@ -2933,21 +2989,33 @@ class IPCHandlers {
if (!realPath) return { success: false, error: "File path not allowed" };
filePath = realPath;
- const diarOpts = {
- numSpeakers: Math.min(
- MAX_SPEAKER_COUNT,
- Math.max(-1, Math.round(Number(options.numSpeakers) || -1))
- ),
- threshold: Math.min(1, Math.max(0, Number(options.threshold) || 0.55)),
- };
+ const numSpeakers = Math.min(
+ MAX_SPEAKER_COUNT,
+ Math.max(-1, Math.round(Number(options.numSpeakers) || -1))
+ );
const { convertToWav } = require("./ffmpegUtils");
const { getSafeTempDir } = require("./safeTempDir");
+ const { resolveClusterThreshold, dropNegligibleClusters } = require("./diarizationPolicy");
+ const { PCM16_MONO_16K_BYTES_PER_SECOND } = require("./transcriptionTimeout");
const wavPath = path.join(getSafeTempDir(), `ow-diarize-${Date.now()}.wav`);
try {
await convertToWav(filePath, wavPath, { sampleRate: 16000, channels: 1 });
- const segments = await this.diarizationManager.diarize(wavPath, diarOpts);
+ // Auto-clustering over-splits long single-mic audio at the 0.55
+ // default, so the threshold ramps with duration unless pinned.
+ const durationSeconds = fs.statSync(wavPath).size / PCM16_MONO_16K_BYTES_PER_SECOND;
+ const threshold = resolveClusterThreshold(durationSeconds, options.threshold);
+
+ let segments = await this.diarizationManager.diarize(wavPath, { numSpeakers, threshold });
+ // The meeting path caps clusters via its expectation resolver; this
+ // path fed raw sherpa output to the merge, which is how a 2-person
+ // voice memo surfaced 46 speakers.
+ segments = dropNegligibleClusters(segments);
+ segments = this.diarizationManager.capSpeakerClusters(
+ segments,
+ numSpeakers > 0 ? numSpeakers : MAX_SPEAKER_COUNT
+ );
return { success: true, segments };
} finally {
try {
@@ -4830,7 +4898,7 @@ class IPCHandlers {
const url = new URL(`${apiUrl}/api/transcribe`);
const data = await postMultipart(url, body, boundary, withPolicyHeaders(authHeader), {
signal: AbortSignal.timeout(CLOUD_UPLOAD_TIMEOUT_MS),
- session: getCloudUploadSession(),
+ session: getInlineCloudUploadSession(),
});
debugLogger.debug(
@@ -4983,7 +5051,7 @@ class IPCHandlers {
withPolicyHeaders(authHeader),
{
signal: AbortSignal.timeout(CLOUD_UPLOAD_TIMEOUT_MS),
- session: getCloudUploadSession(),
+ session: getInlineCloudUploadSession(),
}
);
const responseData = interpretTranscribeResponse(data);
@@ -7935,7 +8003,7 @@ class IPCHandlers {
fileSize,
filePath: path.basename(realCloud),
});
- const { text, warning } = await chunkedCloudTranscribe({
+ const { text, warning, failedChunks, totalChunks } = await chunkedCloudTranscribe({
filePath: realCloud,
apiUrl,
policyHeaders: withPolicyHeaders(authHeader),
@@ -7943,7 +8011,11 @@ class IPCHandlers {
onProgress: (payload) => event.sender.send("upload-transcription-progress", payload),
signal: controller.signal,
});
- return { success: true, text, ...(warning ? { warning } : {}) };
+ return {
+ success: true,
+ text,
+ ...(warning ? { warning, failedChunks, totalChunks } : {}),
+ };
}
const audioBuffer = fs.readFileSync(realCloud);
@@ -7963,7 +8035,7 @@ class IPCHandlers {
controller.signal,
AbortSignal.timeout(CLOUD_UPLOAD_TIMEOUT_MS),
]),
- session: getCloudUploadSession(),
+ session: getInlineCloudUploadSession(),
});
const result = interpretTranscribeResponse(data);
diff --git a/src/hooks/useNoteDragAndDrop.ts b/src/hooks/useNoteDragAndDrop.ts
index 3830c14c3..80b04a259 100644
--- a/src/hooks/useNoteDragAndDrop.ts
+++ b/src/hooks/useNoteDragAndDrop.ts
@@ -231,7 +231,14 @@ export function useNoteDragAndDrop({
},
};
},
- [onCrossSpaceDrop, canCrossSpaceDrop, onCrossSpaceVeto, onHoverTarget, commitDrop, clearHoverTimeout]
+ [
+ onCrossSpaceDrop,
+ canCrossSpaceDrop,
+ onCrossSpaceVeto,
+ onHoverTarget,
+ commitDrop,
+ clearHoverTimeout,
+ ]
);
return { dragState, noteDragHandlers, dropTargetHandlers };
diff --git a/src/locales/de/translation.json b/src/locales/de/translation.json
index 261497325..6e4f5fb06 100644
--- a/src/locales/de/translation.json
+++ b/src/locales/de/translation.json
@@ -601,6 +601,7 @@
"offline": "Du bist offline. Die Cloud-Transkription benötigt eine Internetverbindung.",
"dailyLimitReached": "Du hast dein heutiges Transkriptionslimit erreicht. Versuch es morgen erneut oder upgrade deinen Plan.",
"providerRateLimited": "Dein Transkriptionsanbieter hat die Anfrage begrenzt. Warte kurz und versuch es erneut.",
+ "chunkLoss": "Ein Netzwerkproblem hat zu viel von der Aufnahme verloren. Transkribieren Sie sie über den Verlauf erneut.",
"agentReasoningFailed": "Ihr Sprachagent konnte den Befehl nicht verarbeiten – stattdessen wurde das unbearbeitete Transkript eingefügt.",
"screenContextSkipped": "Der Screenshot konnte nicht gesendet werden, daher lief Ihr Befehl ohne Bildschirmkontext."
},
@@ -2716,6 +2717,8 @@
"cancelTranscription": "Transkription abbrechen",
"transcriptionComplete": "Transkription abgeschlossen",
"partialWarning": "Ein Teil der Audiodatei konnte nicht transkribiert werden.",
+ "partialWarningCount": "{{failed}} von {{total}} Audioabschnitten konnten nicht transkribiert werden.",
+ "chunkLossExceeded": "Wegen eines Netzwerkproblems ging zu viel Audio verloren. Prüfen Sie Ihre Verbindung und versuchen Sie es erneut.",
"openNote": "Notiz öffnen",
"uploadAnother": "Weitere hochladen",
"retry": "Erneut versuchen",
@@ -2762,6 +2765,8 @@
"speakerDetection": "Sprechererkennung",
"speakerDetectionDescription": "Verschiedene Sprecher im Audio identifizieren",
"numSpeakersPlaceholder": "Automatisch erkennen",
+ "numSpeakersLabel": "Anzahl der Sprecher",
+ "numSpeakersHint": "Die genaue Anzahl verbessert die Genauigkeit, besonders bei langen Aufnahmen.",
"downloadingModels": "Modelle werden heruntergeladen...",
"openaiDiarizeNote": "Überschreibt Ihr Transkriptionsmodell mit gpt-4o-transcribe-diarize, dem einzigen OpenAI-Modell mit Sprechererkennung. Die Kosten können von Ihrem gewählten Modell abweichen.",
"mistralDiarizeNote": "Sprechererkennung ohne Zusatzkosten mit Voxtral enthalten.",
@@ -3412,7 +3417,8 @@
"speaker": {
"you": "Du",
"others": "Andere"
- }
+ },
+ "missingAudio": "[fehlendes Audio {{from}}-{{to}}]"
},
"dictationTranslation": {
"title": "Diktat-Übersetzung",
diff --git a/src/locales/en/translation.json b/src/locales/en/translation.json
index 989c39c9e..c6198b4d8 100644
--- a/src/locales/en/translation.json
+++ b/src/locales/en/translation.json
@@ -601,6 +601,7 @@
"offline": "You're offline. Cloud transcription requires an internet connection.",
"dailyLimitReached": "You've reached today's transcription limit. Try again tomorrow or upgrade your plan.",
"providerRateLimited": "Your transcription provider rate-limited the request. Wait a moment and try again.",
+ "chunkLoss": "A network problem lost too much of the recording. Re-transcribe it from History.",
"agentReasoningFailed": "Your voice agent couldn't process the command — the raw transcript was inserted instead.",
"screenContextSkipped": "The screenshot couldn't be sent, so your command ran without screen context."
},
@@ -2716,6 +2717,8 @@
"cancelTranscription": "Cancel transcription",
"transcriptionComplete": "Transcription complete",
"partialWarning": "Part of the audio couldn't be transcribed.",
+ "partialWarningCount": "{{failed}} of {{total}} audio sections couldn't be transcribed.",
+ "chunkLossExceeded": "Too much of the audio was lost to a network problem. Check your connection and try again.",
"openNote": "Open Note",
"uploadAnother": "Upload Another",
"retry": "Retry",
@@ -2762,6 +2765,8 @@
"speakerDetection": "Speaker detection",
"speakerDetectionDescription": "Identify different speakers in the audio",
"numSpeakersPlaceholder": "Auto-detect",
+ "numSpeakersLabel": "Number of speakers",
+ "numSpeakersHint": "Setting the exact count improves accuracy, especially on long recordings.",
"downloadingModels": "Downloading models...",
"openaiDiarizeNote": "Overrides your transcription model with gpt-4o-transcribe-diarize, the only OpenAI model with speaker detection. Costs may differ from your selected model.",
"mistralDiarizeNote": "Speaker detection included at no extra cost with Voxtral.",
@@ -3412,7 +3417,8 @@
"speaker": {
"you": "You",
"others": "Others"
- }
+ },
+ "missingAudio": "[missing audio {{from}}-{{to}}]"
},
"dictationTranslation": {
"title": "Dictation Translation",
diff --git a/src/locales/es/translation.json b/src/locales/es/translation.json
index 26861bb4b..5db2d9320 100644
--- a/src/locales/es/translation.json
+++ b/src/locales/es/translation.json
@@ -601,6 +601,7 @@
"offline": "Estás sin conexión. La transcripción en la nube requiere conexión a Internet.",
"dailyLimitReached": "Has alcanzado el límite de transcripción de hoy. Inténtalo mañana o mejora tu plan.",
"providerRateLimited": "Tu proveedor de transcripción limitó la solicitud. Espera un momento e inténtalo de nuevo.",
+ "chunkLoss": "Un problema de red hizo que se perdiera demasiada parte de la grabación. Vuelve a transcribirla desde el historial.",
"agentReasoningFailed": "Tu agente de voz no pudo procesar el comando; se insertó la transcripción sin procesar.",
"screenContextSkipped": "No se pudo enviar la captura, así que tu comando se ejecutó sin contexto de pantalla."
},
@@ -2573,6 +2574,8 @@
"cancelTranscription": "Cancelar transcripción",
"transcriptionComplete": "Transcripción completada",
"partialWarning": "No se pudo transcribir una parte del audio.",
+ "partialWarningCount": "No se pudieron transcribir {{failed}} de {{total}} secciones del audio.",
+ "chunkLossExceeded": "Se perdió demasiado audio por un problema de red. Comprueba tu conexión e inténtalo de nuevo.",
"openNote": "Abrir nota",
"uploadAnother": "Subir otro",
"retry": "Reintentar",
@@ -2619,6 +2622,8 @@
"speakerDetection": "Detección de hablantes",
"speakerDetectionDescription": "Identificar diferentes hablantes en el audio",
"numSpeakersPlaceholder": "Detección automática",
+ "numSpeakersLabel": "Número de hablantes",
+ "numSpeakersHint": "Indicar el número exacto mejora la precisión, sobre todo en grabaciones largas.",
"downloadingModels": "Descargando modelos...",
"openaiDiarizeNote": "Reemplaza tu modelo de transcripción con gpt-4o-transcribe-diarize, el único modelo de OpenAI con detección de hablantes. Los costos pueden diferir de tu modelo seleccionado.",
"mistralDiarizeNote": "Detección de hablantes incluida sin costo adicional con Voxtral.",
@@ -3352,7 +3357,8 @@
"speaker": {
"you": "Tú",
"others": "Otros"
- }
+ },
+ "missingAudio": "[audio faltante {{from}}-{{to}}]"
},
"workspaces": {
"switcher": {
diff --git a/src/locales/fr/translation.json b/src/locales/fr/translation.json
index 498fc53ae..87287370e 100644
--- a/src/locales/fr/translation.json
+++ b/src/locales/fr/translation.json
@@ -601,6 +601,7 @@
"offline": "Vous êtes hors ligne. La transcription cloud nécessite une connexion Internet.",
"dailyLimitReached": "Vous avez atteint votre limite de transcription du jour. Réessayez demain ou améliorez votre forfait.",
"providerRateLimited": "Votre fournisseur de transcription a limité la requête. Patientez un instant et réessayez.",
+ "chunkLoss": "Un problème réseau a fait perdre une trop grande partie de l'enregistrement. Retranscrivez-le depuis l'historique.",
"agentReasoningFailed": "Votre agent vocal n'a pas pu traiter la commande ; la transcription brute a été insérée à la place.",
"screenContextSkipped": "La capture n'a pas pu être envoyée ; votre commande a été exécutée sans contexte d'écran."
},
@@ -2716,6 +2717,8 @@
"cancelTranscription": "Annuler la transcription",
"transcriptionComplete": "Transcription terminée",
"partialWarning": "Une partie de l'audio n'a pas pu être transcrite.",
+ "partialWarningCount": "{{failed}} sections audio sur {{total}} n'ont pas pu être transcrites.",
+ "chunkLossExceeded": "Une trop grande partie de l'audio a été perdue à cause d'un problème réseau. Vérifiez votre connexion et réessayez.",
"openNote": "Ouvrir la note",
"uploadAnother": "Téléverser un autre fichier",
"retry": "Réessayer",
@@ -2762,6 +2765,8 @@
"speakerDetection": "Détection des intervenants",
"speakerDetectionDescription": "Identifier les différents intervenants dans l'audio",
"numSpeakersPlaceholder": "Détection automatique",
+ "numSpeakersLabel": "Nombre d'intervenants",
+ "numSpeakersHint": "Indiquer le nombre exact améliore la précision, surtout pour les enregistrements longs.",
"downloadingModels": "Téléchargement des modèles...",
"openaiDiarizeNote": "Remplace votre modèle de transcription par gpt-4o-transcribe-diarize, le seul modèle OpenAI avec détection des locuteurs. Les coûts peuvent différer de votre modèle sélectionné.",
"mistralDiarizeNote": "Détection des locuteurs incluse sans frais supplémentaires avec Voxtral.",
@@ -3412,7 +3417,8 @@
"speaker": {
"you": "Vous",
"others": "Autres"
- }
+ },
+ "missingAudio": "[audio manquant {{from}}-{{to}}]"
},
"dictationTranslation": {
"title": "Traduction de Dictée",
diff --git a/src/locales/it/translation.json b/src/locales/it/translation.json
index e35283feb..3b1086773 100644
--- a/src/locales/it/translation.json
+++ b/src/locales/it/translation.json
@@ -601,6 +601,7 @@
"offline": "Sei offline. La trascrizione cloud richiede una connessione Internet.",
"dailyLimitReached": "Hai raggiunto il limite di trascrizioni di oggi. Riprova domani o aggiorna il tuo piano.",
"providerRateLimited": "Il tuo provider di trascrizione ha limitato la richiesta. Attendi un momento e riprova.",
+ "chunkLoss": "Un problema di rete ha fatto perdere troppa parte della registrazione. Ritrascrivila dalla cronologia.",
"agentReasoningFailed": "L'agente vocale non è riuscito a elaborare il comando; è stata inserita la trascrizione grezza.",
"screenContextSkipped": "Non è stato possibile inviare lo screenshot, quindi il comando è stato eseguito senza contesto dello schermo."
},
@@ -2525,6 +2526,8 @@
"cancelTranscription": "Annulla trascrizione",
"transcriptionComplete": "Trascrizione completata",
"partialWarning": "Non è stato possibile trascrivere una parte dell'audio.",
+ "partialWarningCount": "Non è stato possibile trascrivere {{failed}} sezioni audio su {{total}}.",
+ "chunkLossExceeded": "Troppo audio è andato perso a causa di un problema di rete. Controlla la connessione e riprova.",
"openNote": "Apri nota",
"uploadAnother": "Carica un altro",
"retry": "Riprova",
@@ -2571,6 +2574,8 @@
"speakerDetection": "Rilevamento parlanti",
"speakerDetectionDescription": "Identifica i diversi parlanti nell'audio",
"numSpeakersPlaceholder": "Rilevamento automatico",
+ "numSpeakersLabel": "Numero di parlanti",
+ "numSpeakersHint": "Indicare il numero esatto migliora la precisione, soprattutto nelle registrazioni lunghe.",
"downloadingModels": "Download modelli in corso...",
"openaiDiarizeNote": "Sostituisce il tuo modello di trascrizione con gpt-4o-transcribe-diarize, l’unico modello OpenAI con rilevamento dei parlanti. I costi potrebbero differire dal modello selezionato.",
"mistralDiarizeNote": "Rilevamento dei parlanti incluso senza costi aggiuntivi con Voxtral.",
@@ -3352,7 +3357,8 @@
"speaker": {
"you": "Tu",
"others": "Altri"
- }
+ },
+ "missingAudio": "[audio mancante {{from}}-{{to}}]"
},
"workspaces": {
"switcher": {
diff --git a/src/locales/ja/translation.json b/src/locales/ja/translation.json
index be2d39ced..7afc611ea 100644
--- a/src/locales/ja/translation.json
+++ b/src/locales/ja/translation.json
@@ -601,6 +601,7 @@
"offline": "オフラインです。クラウド文字起こしにはインターネット接続が必要です。",
"dailyLimitReached": "本日の文字起こし上限に達しました。明日もう一度お試しいただくか、プランをアップグレードしてください。",
"providerRateLimited": "文字起こしプロバイダーがリクエストを制限しました。しばらく待ってからもう一度お試しください。",
+ "chunkLoss": "ネットワークの問題で録音の大部分が失われました。履歴から再文字起こししてください。",
"agentReasoningFailed": "ボイスエージェントがコマンドを処理できなかったため、未処理の文字起こしを挿入しました。",
"screenContextSkipped": "スクリーンショットを送信できなかったため、画面コンテキストなしでコマンドを実行しました。"
},
@@ -2509,6 +2510,8 @@
"cancelTranscription": "文字起こしをキャンセル",
"transcriptionComplete": "文字起こし完了",
"partialWarning": "音声の一部を文字起こしできませんでした。",
+ "partialWarningCount": "{{total}}個の音声セクションのうち{{failed}}個を文字起こしできませんでした。",
+ "chunkLossExceeded": "ネットワークの問題で音声の大部分が失われました。接続を確認してもう一度お試しください。",
"openNote": "ノートを開く",
"uploadAnother": "別のファイルをアップロード",
"retry": "再試行",
@@ -2555,6 +2558,8 @@
"speakerDetection": "話者検出",
"speakerDetectionDescription": "音声内の異なる話者を識別",
"numSpeakersPlaceholder": "自動検出",
+ "numSpeakersLabel": "話者の人数",
+ "numSpeakersHint": "正確な人数を設定すると、特に長い録音で精度が向上します。",
"downloadingModels": "モデルをダウンロード中...",
"openaiDiarizeNote": "話者検出をサポートする唯一のOpenAIモデルであるgpt-4o-transcribe-diarizeに切り替わります。選択したモデルとコストが異なる場合があります。",
"mistralDiarizeNote": "Voxtralでの話者検出は追加料金なしで利用できます。",
@@ -3352,7 +3357,8 @@
"speaker": {
"you": "あなた",
"others": "他の人"
- }
+ },
+ "missingAudio": "[音声欠落 {{from}}-{{to}}]"
},
"workspaces": {
"switcher": {
diff --git a/src/locales/pt/translation.json b/src/locales/pt/translation.json
index 345e65fe4..993a28801 100644
--- a/src/locales/pt/translation.json
+++ b/src/locales/pt/translation.json
@@ -580,6 +580,7 @@
"offline": "Você está offline. A transcrição na nuvem exige conexão com a internet.",
"dailyLimitReached": "Você atingiu o limite de transcrições de hoje. Tente novamente amanhã ou faça upgrade do seu plano.",
"providerRateLimited": "Seu provedor de transcrição limitou a solicitação. Aguarde um momento e tente novamente.",
+ "chunkLoss": "Um problema de rede fez perder grande parte da gravação. Retranscreva-a a partir do histórico.",
"agentReasoningFailed": "Seu agente de voz não conseguiu processar o comando; a transcrição bruta foi inserida no lugar.",
"screenContextSkipped": "Não foi possível enviar a captura, então seu comando foi executado sem contexto de tela."
},
@@ -2525,6 +2526,8 @@
"cancelTranscription": "Cancelar transcrição",
"transcriptionComplete": "Transcrição concluída",
"partialWarning": "Não foi possível transcrever parte do áudio.",
+ "partialWarningCount": "Não foi possível transcrever {{failed}} de {{total}} seções do áudio.",
+ "chunkLossExceeded": "Muito áudio foi perdido devido a um problema de rede. Verifique sua conexão e tente novamente.",
"openNote": "Abrir nota",
"uploadAnother": "Enviar outro",
"retry": "Tentar novamente",
@@ -2571,6 +2574,8 @@
"speakerDetection": "Detecção de locutores",
"speakerDetectionDescription": "Identificar diferentes locutores no áudio",
"numSpeakersPlaceholder": "Detecção automática",
+ "numSpeakersLabel": "Número de falantes",
+ "numSpeakersHint": "Definir o número exato melhora a precisão, principalmente em gravações longas.",
"downloadingModels": "Baixando modelos...",
"openaiDiarizeNote": "Substitui seu modelo de transcrição por gpt-4o-transcribe-diarize, o único modelo OpenAI com detecção de falantes. Os custos podem diferir do modelo selecionado.",
"mistralDiarizeNote": "Detecção de falantes incluída sem custo adicional com Voxtral.",
@@ -3352,7 +3357,8 @@
"speaker": {
"you": "Você",
"others": "Outros"
- }
+ },
+ "missingAudio": "[áudio ausente {{from}}-{{to}}]"
},
"workspaces": {
"switcher": {
diff --git a/src/locales/ru/translation.json b/src/locales/ru/translation.json
index db0d8579a..a301b2a95 100644
--- a/src/locales/ru/translation.json
+++ b/src/locales/ru/translation.json
@@ -601,6 +601,7 @@
"offline": "Вы офлайн. Облачная транскрипция требует подключения к интернету.",
"dailyLimitReached": "Достигнут дневной лимит транскрипций. Попробуйте завтра или обновите тариф.",
"providerRateLimited": "Ваш провайдер транскрипции ограничил запрос. Подождите немного и попробуйте снова.",
+ "chunkLoss": "Из-за проблемы с сетью потеряна значительная часть записи. Повторите расшифровку из истории.",
"agentReasoningFailed": "Голосовой агент не смог обработать команду — вместо этого вставлена необработанная расшифровка.",
"screenContextSkipped": "Снимок экрана не удалось отправить, поэтому команда выполнена без контекста экрана."
},
@@ -2547,6 +2548,8 @@
"cancelTranscription": "Отменить транскрибирование",
"transcriptionComplete": "Транскрибирование завершено",
"partialWarning": "Не удалось транскрибировать часть аудио.",
+ "partialWarningCount": "Не удалось транскрибировать {{failed}} из {{total}} фрагментов аудио.",
+ "chunkLossExceeded": "Из-за проблемы с сетью потеряно слишком много аудио. Проверьте подключение и попробуйте снова.",
"openNote": "Открыть заметку",
"uploadAnother": "Загрузить ещё",
"retry": "Повторить",
@@ -2593,6 +2596,8 @@
"speakerDetection": "Определение говорящих",
"speakerDetectionDescription": "Определить разных говорящих в аудио",
"numSpeakersPlaceholder": "Автоопределение",
+ "numSpeakersLabel": "Количество говорящих",
+ "numSpeakersHint": "Точное число повышает точность, особенно для длинных записей.",
"downloadingModels": "Загрузка моделей...",
"openaiDiarizeNote": "Заменяет вашу модель транскрипции на gpt-4o-transcribe-diarize — единственную модель OpenAI с определением говорящих. Стоимость может отличаться от выбранной модели.",
"mistralDiarizeNote": "Определение говорящих включено без дополнительной оплаты с Voxtral.",
@@ -3382,7 +3387,8 @@
"speaker": {
"you": "Вы",
"others": "Другие"
- }
+ },
+ "missingAudio": "[аудио отсутствует {{from}}-{{to}}]"
},
"workspaces": {
"switcher": {
diff --git a/src/locales/zh-CN/translation.json b/src/locales/zh-CN/translation.json
index 0f01c5d7c..981ca3140 100644
--- a/src/locales/zh-CN/translation.json
+++ b/src/locales/zh-CN/translation.json
@@ -601,6 +601,7 @@
"offline": "你目前离线。云端转录需要互联网连接。",
"dailyLimitReached": "已达今日转录上限。请明天再试或升级你的套餐。",
"providerRateLimited": "你的转录服务商限制了请求频率。请稍候再试。",
+ "chunkLoss": "网络问题导致录音大部分丢失。请从历史记录中重新转录。",
"agentReasoningFailed": "语音助手无法处理该命令,已改为插入原始转录文本。",
"screenContextSkipped": "截图无法发送,因此已在没有屏幕上下文的情况下执行你的命令。"
},
@@ -2525,6 +2526,8 @@
"cancelTranscription": "取消转录",
"transcriptionComplete": "转录完成",
"partialWarning": "部分音频内容无法转录。",
+ "partialWarningCount": "{{total}}个音频片段中有{{failed}}个无法转写。",
+ "chunkLossExceeded": "由于网络问题丢失了过多音频。请检查网络连接后重试。",
"openNote": "打开笔记",
"uploadAnother": "上传另一个",
"retry": "重试",
@@ -2571,6 +2574,8 @@
"speakerDetection": "说话人检测",
"speakerDetectionDescription": "识别音频中的不同说话人",
"numSpeakersPlaceholder": "自动检测",
+ "numSpeakersLabel": "说话人数",
+ "numSpeakersHint": "设置准确的人数可以提高准确性,尤其是长录音。",
"downloadingModels": "正在下载模型...",
"openaiDiarizeNote": "将替换您的转录模型为 gpt-4o-transcribe-diarize,这是唯一支持说话人检测的 OpenAI 模型。费用可能与您选择的模型不同。",
"mistralDiarizeNote": "使用 Voxtral 时说话人检测无需额外费用。",
@@ -3352,7 +3357,8 @@
"speaker": {
"you": "你",
"others": "其他人"
- }
+ },
+ "missingAudio": "[音频缺失 {{from}}-{{to}}]"
},
"workspaces": {
"switcher": {
diff --git a/src/locales/zh-TW/translation.json b/src/locales/zh-TW/translation.json
index a742f51fa..b6fb8ef1e 100644
--- a/src/locales/zh-TW/translation.json
+++ b/src/locales/zh-TW/translation.json
@@ -601,6 +601,7 @@
"offline": "你目前離線。雲端轉錄需要網際網路連線。",
"dailyLimitReached": "已達今日的轉錄上限。請明天再試或升級你的方案。",
"providerRateLimited": "你的轉錄服務商限制了請求頻率。請稍候再試。",
+ "chunkLoss": "網路問題導致錄音大部分遺失。請從歷史記錄中重新轉錄。",
"agentReasoningFailed": "語音助理無法處理該指令,已改為插入原始逐字稿。",
"screenContextSkipped": "截圖無法傳送,因此已在沒有螢幕內容的情況下執行你的指令。"
},
@@ -2525,6 +2526,8 @@
"cancelTranscription": "取消轉錄",
"transcriptionComplete": "轉錄完成",
"partialWarning": "部分音訊內容無法轉錄。",
+ "partialWarningCount": "{{total}}個音訊片段中有{{failed}}個無法轉錄。",
+ "chunkLossExceeded": "由於網路問題遺失了過多音訊。請檢查網路連線後重試。",
"openNote": "開啟筆記",
"uploadAnother": "上傳另一個",
"retry": "重試",
@@ -2571,6 +2574,8 @@
"speakerDetection": "說話者偵測",
"speakerDetectionDescription": "辨識音訊中的不同說話者",
"numSpeakersPlaceholder": "自動偵測",
+ "numSpeakersLabel": "說話人數",
+ "numSpeakersHint": "設定準確的人數可以提高準確度,尤其是長錄音。",
"downloadingModels": "正在下載模型...",
"openaiDiarizeNote": "將替換您的轉錄模型為 gpt-4o-transcribe-diarize,這是唯一支援說話人偵測的 OpenAI 模型。費用可能與您選擇的模型不同。",
"mistralDiarizeNote": "使用 Voxtral 時說話人偵測無需額外費用。",
@@ -3352,7 +3357,8 @@
"speaker": {
"you": "你",
"others": "其他人"
- }
+ },
+ "missingAudio": "[音訊缺失 {{from}}-{{to}}]"
},
"workspaces": {
"switcher": {
diff --git a/src/services/fileTranscription.ts b/src/services/fileTranscription.ts
index 75fb63abd..a2a775f6a 100644
--- a/src/services/fileTranscription.ts
+++ b/src/services/fileTranscription.ts
@@ -14,6 +14,9 @@ export interface FileTranscriptionResult {
code?: string;
diarized?: boolean;
warning?: string;
+ // Set alongside `warning` by the chunked cloud path: how much audio was lost.
+ failedChunks?: number;
+ totalChunks?: number;
}
export interface DiarizationSettings {
diff --git a/src/stores/batchQueueStore.ts b/src/stores/batchQueueStore.ts
index ca2466ab8..38871dde8 100644
--- a/src/stores/batchQueueStore.ts
+++ b/src/stores/batchQueueStore.ts
@@ -1,7 +1,7 @@
import { create } from "zustand";
import { transcribeFileWithSpeakers } from "../services/fileTranscription";
import type { FileTranscriptionConfig, DiarizationSettings } from "../services/fileTranscription";
-import { DOWNLOAD_ERROR_KEYS } from "../components/notes/shared";
+import { DOWNLOAD_ERROR_KEYS, transcriptionErrorKey } from "../components/notes/shared";
import { getSettings } from "./settingsStore";
import { isTranscriptionContextAllowed } from "./policyRules";
import { usePolicyStore } from "./policyStore";
@@ -208,9 +208,9 @@ export function processBatchQueue(
updateItem(item.id, {
status: "error",
error:
- transcriptionResult.code === "NO_SPEECH_DETECTED"
- ? "noSpeechDetected"
- : transcriptionResult.error || "batchTranscriptionFailed",
+ transcriptionErrorKey(transcriptionResult) ||
+ transcriptionResult.error ||
+ "batchTranscriptionFailed",
});
return;
}
@@ -251,7 +251,8 @@ export function processBatchQueue(
} catch (err) {
updateItem(item.id, {
status: "error",
- error: err instanceof Error ? err.message : "batchUnknownError",
+ error:
+ transcriptionErrorKey(err) || (err instanceof Error ? err.message : "batchUnknownError"),
});
} finally {
// Nothing else owns the temp file, so delete it even for a stale run.
diff --git a/src/types/electron.ts b/src/types/electron.ts
index 957a16faa..bf7e2bf2d 100644
--- a/src/types/electron.ts
+++ b/src/types/electron.ts
@@ -1791,6 +1791,8 @@ declare global {
success: boolean;
text?: string;
warning?: string;
+ failedChunks?: number;
+ totalChunks?: number;
} & PolicyFailureMetadata
>;
diff --git a/test/helpers/cloudChunkPolicy.test.js b/test/helpers/cloudChunkPolicy.test.js
index d1810b1b3..8d6101aaa 100644
--- a/test/helpers/cloudChunkPolicy.test.js
+++ b/test/helpers/cloudChunkPolicy.test.js
@@ -6,14 +6,22 @@ const {
CLOUD_CHUNK_MAX_ATTEMPTS,
CLOUD_CHUNK_GLOBAL_CONCURRENCY,
CLOUD_POOL_TEARDOWN_COOLDOWN_MS,
+ CLOUD_CHUNK_MAX_TEARDOWN_REFUNDS,
+ CLOUD_CHUNK_MAX_LOSS_RATIO,
+ SILENT_CHUNK,
FATAL_CHUNK_CODES,
isTransientChunkError,
isNetworkLevelFailure,
+ isConnectionPoisoningFailure,
+ isTeardownCollateral,
+ summarizeChunkResults,
+ assembleChunkTranscript,
chunkRetryDelayMs,
abortableSleep,
createTeardownGate,
createUploadSlots,
} = require("../../src/helpers/cloudChunkPolicy");
+const { changeLanguage } = require("../../src/helpers/i18nMain");
const { createAbortError } = require("../../src/helpers/abortError");
// The numbers are the #1326 contract: a dead upload fails within 2 minutes and
@@ -43,7 +51,7 @@ test("teardown gate admits one drop per cooldown window", () => {
assert.equal(gate(), true, "a later wedge must still be recoverable");
});
-test("auth and quota doom the job; silence dooms only its own chunk", () => {
+test("auth and quota doom the job; a no-speech chunk is nonfatal", () => {
assert.equal(FATAL_CHUNK_CODES.has("AUTH_EXPIRED"), true);
assert.equal(FATAL_CHUNK_CODES.has("LIMIT_REACHED"), true);
// Silence in one segment says nothing about the rest, so the siblings must
@@ -122,6 +130,140 @@ test("abortableSleep rejects promptly when aborted mid-wait", async () => {
assert.ok(Date.now() - start < 5_000, "abort did not interrupt the sleep");
});
+// A fatal TLS alert (Nat Molina's net::ERR_SSL_BAD_RECORD_MAC_ALERT) kills the
+// HTTP/2 connection and every stream on it; waiting out the 30s cooldown just
+// feeds retries back into the same sick pool until their budgets are gone.
+test("fatal TLS/protocol errors are recognized as connection-poisoning", () => {
+ for (const msg of [
+ "net::ERR_SSL_BAD_RECORD_MAC_ALERT",
+ "Failed to fetch: net::ERR_SSL_PROTOCOL_ERROR",
+ "net::ERR_HTTP2_PROTOCOL_ERROR",
+ "net::ERR_QUIC_PROTOCOL_ERROR",
+ ]) {
+ assert.equal(isConnectionPoisoningFailure(new Error(msg)), true, msg);
+ }
+});
+
+test("ordinary failures are not connection-poisoning", () => {
+ assert.equal(isConnectionPoisoningFailure(new Error("net::ERR_CONNECTION_CLOSED")), false);
+ assert.equal(isConnectionPoisoningFailure(new Error("aborted")), false);
+ assert.equal(
+ isConnectionPoisoningFailure(Object.assign(new Error("x"), { statusCode: 502 })),
+ false
+ );
+ assert.equal(isConnectionPoisoningFailure({}), false);
+ assert.equal(isConnectionPoisoningFailure(null), false);
+});
+
+test("a forced teardown bypasses the cooldown and restarts the window", () => {
+ let now = 0;
+ const gate = createTeardownGate(CLOUD_POOL_TEARDOWN_COOLDOWN_MS, () => now);
+
+ assert.equal(gate(), true);
+ assert.equal(gate(true), true, "poisoned connection must be droppable inside the cooldown");
+ now += CLOUD_POOL_TEARDOWN_COOLDOWN_MS - 1;
+ assert.equal(gate(), false, "the forced drop restarts the cooldown window");
+ now += 1;
+ assert.equal(gate(), true);
+});
+
+// closeAllConnections() aborts every sibling still on the wire. Those failures
+// are self-inflicted — the sibling's connection wasn't sick — so their attempts
+// are refunded instead of charged.
+test("teardown collateral: network-level failure during a teardown is refunded", () => {
+ const abortLike = new Error("net::ERR_ABORTED");
+ assert.equal(
+ isTeardownCollateral(abortLike, { timedOut: false, teardownsDuringAttempt: 1 }),
+ true
+ );
+});
+
+test("teardown collateral requires a teardown during the attempt", () => {
+ const abortLike = new Error("net::ERR_ABORTED");
+ assert.equal(
+ isTeardownCollateral(abortLike, { timedOut: false, teardownsDuringAttempt: 0 }),
+ false
+ );
+});
+
+test("a timeout or an answered request is never teardown collateral", () => {
+ assert.equal(
+ isTeardownCollateral(new Error("aborted"), { timedOut: true, teardownsDuringAttempt: 1 }),
+ false,
+ "a stalled upload is the chunk's own failure"
+ );
+ assert.equal(
+ isTeardownCollateral(Object.assign(new Error("x"), { statusCode: 502 }), {
+ timedOut: false,
+ teardownsDuringAttempt: 1,
+ }),
+ false,
+ "an HTTP answer proves the request outlived the teardown"
+ );
+ assert.equal(
+ isTeardownCollateral(Object.assign(new Error("x"), { code: "AUTH_EXPIRED" }), {
+ timedOut: false,
+ teardownsDuringAttempt: 1,
+ }),
+ false
+ );
+});
+
+test("refunds are bounded and the loss ceiling is a real fraction", () => {
+ assert.ok(
+ Number.isInteger(CLOUD_CHUNK_MAX_TEARDOWN_REFUNDS) && CLOUD_CHUNK_MAX_TEARDOWN_REFUNDS > 0
+ );
+ assert.ok(CLOUD_CHUNK_MAX_LOSS_RATIO > 0 && CLOUD_CHUNK_MAX_LOSS_RATIO < 1);
+});
+
+// Survivors used to be joined with a bare space — a 16-minute fragment of a
+// 73-minute file read as a seamless transcript.
+test("assembleChunkTranscript marks failed spans and collapses consecutive holes", () => {
+ const results = [{ text: "one" }, null, null, { text: "four" }, null];
+ assert.equal(
+ assembleChunkTranscript(results, 240),
+ "one [missing audio 4:00-12:00] four [missing audio 16:00-20:00]"
+ );
+});
+
+test("assembleChunkTranscript handles a hole at the start and hour-long offsets", () => {
+ const results = [null, { text: "later" }];
+ assert.equal(assembleChunkTranscript(results, 3600), "[missing audio 0:00-1:00:00] later");
+});
+
+test("assembleChunkTranscript with no holes matches the plain join", () => {
+ const results = [{ text: " a " }, { text: "b\n\nc" }];
+ assert.equal(assembleChunkTranscript(results, 240), "a b c");
+});
+
+test("silent chunks are omitted without consuming the failure budget", () => {
+ const results = [{ text: "one" }, SILENT_CHUNK, null, { text: "four" }];
+ assert.deepEqual(summarizeChunkResults(results), {
+ responses: [{ text: "one" }, { text: "four" }],
+ failedChunks: 1,
+ silentChunks: 1,
+ });
+ assert.equal(assembleChunkTranscript(results, 240), "one [missing audio 8:00-12:00] four");
+});
+
+test("the final missing-audio marker ends at the real input duration", () => {
+ const results = new Array(19).fill(SILENT_CHUNK);
+ results[18] = null;
+ assert.equal(assembleChunkTranscript(results, 240, 73 * 60), "[missing audio 1:12:00-1:13:00]");
+});
+
+// The marker is persisted into notes and pasted by dictation, so it belongs to
+// the UI language like the [Speaker N] labels do.
+test("the missing-audio marker follows the UI language", (t) => {
+ changeLanguage("de");
+ t.after(() => changeLanguage("en"));
+
+ assert.equal(
+ assembleChunkTranscript([{ text: "eins" }, null], 240),
+ "eins [fehlendes Audio 4:00-8:00]"
+ );
+});
+
test("upload slots cap concurrent holders across independent acquirers", async () => {
const slots = createUploadSlots(2);
const r1 = await slots.acquire();
diff --git a/test/helpers/diarizationPolicy.test.js b/test/helpers/diarizationPolicy.test.js
new file mode 100644
index 000000000..0573566aa
--- /dev/null
+++ b/test/helpers/diarizationPolicy.test.js
@@ -0,0 +1,100 @@
+const test = require("node:test");
+const assert = require("node:assert/strict");
+
+const {
+ DEFAULT_CLUSTER_THRESHOLD,
+ LONG_AUDIO_CLUSTER_THRESHOLD,
+ THRESHOLD_RAMP_START_SECONDS,
+ THRESHOLD_RAMP_END_SECONDS,
+ MIN_CLUSTER_TOTAL_SECONDS,
+ clusterThresholdForDuration,
+ resolveClusterThreshold,
+ dropNegligibleClusters,
+} = require("../../src/helpers/diarizationPolicy");
+
+// 0.55 is tuned for short clean audio; on a 73-minute single-mic voice memo one
+// speaker's embeddings spread past it and clustering stopped merging at 46
+// "speakers". The threshold must grow with duration.
+test("short audio keeps the sherpa default threshold", () => {
+ assert.equal(clusterThresholdForDuration(0), DEFAULT_CLUSTER_THRESHOLD);
+ assert.equal(
+ clusterThresholdForDuration(THRESHOLD_RAMP_START_SECONDS),
+ DEFAULT_CLUSTER_THRESHOLD
+ );
+});
+
+test("threshold ramps between the boundaries and is monotonic", () => {
+ const mid = clusterThresholdForDuration(
+ (THRESHOLD_RAMP_START_SECONDS + THRESHOLD_RAMP_END_SECONDS) / 2
+ );
+ assert.ok(mid > DEFAULT_CLUSTER_THRESHOLD && mid < LONG_AUDIO_CLUSTER_THRESHOLD);
+ assert.ok(
+ clusterThresholdForDuration(THRESHOLD_RAMP_START_SECONDS + 60) <= mid,
+ "threshold must not decrease with duration"
+ );
+});
+
+test("hour-plus audio is clamped to the long-audio threshold", () => {
+ assert.equal(
+ clusterThresholdForDuration(THRESHOLD_RAMP_END_SECONDS),
+ LONG_AUDIO_CLUSTER_THRESHOLD
+ );
+ assert.equal(clusterThresholdForDuration(3 * 3600), LONG_AUDIO_CLUSTER_THRESHOLD);
+});
+
+test("unknown duration falls back to the default threshold", () => {
+ assert.equal(clusterThresholdForDuration(NaN), DEFAULT_CLUSTER_THRESHOLD);
+ assert.equal(clusterThresholdForDuration(undefined), DEFAULT_CLUSTER_THRESHOLD);
+ assert.equal(clusterThresholdForDuration(-5), DEFAULT_CLUSTER_THRESHOLD);
+});
+
+test("explicit thresholds accept zero, clamp bounds, and reject non-finite values", () => {
+ assert.equal(resolveClusterThreshold(3600, 0), 0);
+ assert.equal(resolveClusterThreshold(3600, 2), 1);
+ assert.equal(resolveClusterThreshold(3600, -1), 0);
+ assert.equal(resolveClusterThreshold(3600, ""), clusterThresholdForDuration(3600));
+ assert.equal(resolveClusterThreshold(3600, "not-a-number"), clusterThresholdForDuration(3600));
+ assert.equal(resolveClusterThreshold(3600, Infinity), clusterThresholdForDuration(3600));
+});
+
+// Every cluster — however tiny — wins at least one sentence in the character-
+// proportional merge, so a stray blip becomes a phantom [Speaker N].
+test("clusters below the minimum total speaking time are dropped", () => {
+ const segments = [
+ { start: 0, end: 30, speaker: "speaker_0" },
+ { start: 30, end: 30.4, speaker: "speaker_7" },
+ { start: 31, end: 60, speaker: "speaker_1" },
+ { start: 60, end: 60.3, speaker: "speaker_7" },
+ ];
+ const kept = dropNegligibleClusters(segments);
+ assert.deepEqual(
+ kept.map((s) => s.speaker),
+ ["speaker_0", "speaker_1"]
+ );
+});
+
+test("a cluster whose short segments add up past the minimum survives", () => {
+ const segments = [
+ { start: 0, end: 30, speaker: "speaker_0" },
+ { start: 30, end: 30.6, speaker: "speaker_1" },
+ { start: 40, end: 40.6, speaker: "speaker_1" },
+ ];
+ assert.equal(dropNegligibleClusters(segments).length, 3);
+});
+
+test("dropping never empties the result or touches a clean input", () => {
+ const allTiny = [
+ { start: 0, end: 0.3, speaker: "speaker_0" },
+ { start: 1, end: 1.2, speaker: "speaker_1" },
+ ];
+ assert.deepEqual(dropNegligibleClusters(allTiny), allTiny);
+
+ const clean = [{ start: 0, end: 10, speaker: "speaker_0" }];
+ assert.deepEqual(dropNegligibleClusters(clean), clean);
+ assert.deepEqual(dropNegligibleClusters([]), []);
+ assert.equal(dropNegligibleClusters(null), null);
+});
+
+test("minimum speaking time constant is sane", () => {
+ assert.ok(MIN_CLUSTER_TOTAL_SECONDS >= 0.5 && MIN_CLUSTER_TOTAL_SECONDS <= 5);
+});
diff --git a/test/helpers/ffmpegUtils.test.js b/test/helpers/ffmpegUtils.test.js
new file mode 100644
index 000000000..ca6ae2901
--- /dev/null
+++ b/test/helpers/ffmpegUtils.test.js
@@ -0,0 +1,14 @@
+const test = require("node:test");
+const assert = require("node:assert/strict");
+
+const { parseFfmpegDuration } = require("../../src/helpers/ffmpegUtils");
+
+test("parseFfmpegDuration reads the input duration from ffmpeg output", () => {
+ const stderr = "Input #0, mp3, from 'recording.mp3':\n Duration: 01:13:00.25, start: 0.000000";
+ assert.equal(parseFfmpegDuration(stderr), 4380.25);
+});
+
+test("parseFfmpegDuration returns null when ffmpeg reports no duration", () => {
+ assert.equal(parseFfmpegDuration("Duration: N/A"), null);
+ assert.equal(parseFfmpegDuration(""), null);
+});
diff --git a/test/helpers/ipcPolicyHeaderCoverage.test.js b/test/helpers/ipcPolicyHeaderCoverage.test.js
index 900cef0a2..63d41be9c 100644
--- a/test/helpers/ipcPolicyHeaderCoverage.test.js
+++ b/test/helpers/ipcPolicyHeaderCoverage.test.js
@@ -37,6 +37,25 @@ test("all OpenWhispr multipart transcription transports use policy headers", ()
);
});
+test("chunk pool recovery cannot abort inline multipart uploads", () => {
+ const chunkedStart = source.indexOf("async function chunkedCloudTranscribe");
+ const handlersStart = source.indexOf("class IPCHandlers");
+ assert.ok(chunkedStart >= 0 && handlersStart > chunkedStart);
+
+ const chunkedSource = source.slice(chunkedStart, handlersStart);
+ assert.match(chunkedSource, /session:\s*getChunkCloudUploadSession\(\)/);
+ assert.doesNotMatch(chunkedSource, /getInlineCloudUploadSession/);
+ assert.match(
+ source,
+ /async function dropUploadConnections[\s\S]{0,300}?getChunkCloudUploadSession\(\)\.closeAllConnections/
+ );
+
+ const handlerSource = source.slice(handlersStart);
+ assert.match(handlerSource, /session:\s*getInlineCloudUploadSession\(\)/);
+ assert.doesNotMatch(handlerSource, /session:\s*getChunkCloudUploadSession\(\)/);
+ assert.doesNotMatch(source, /session:\s*getCloudUploadSession\(/);
+});
+
test("the shared realtime-token helper applies policy headers", () => {
assert.match(
source,
diff --git a/test/helpers/transcriptionErrorKeys.test.js b/test/helpers/transcriptionErrorKeys.test.js
new file mode 100644
index 000000000..7feec656e
--- /dev/null
+++ b/test/helpers/transcriptionErrorKeys.test.js
@@ -0,0 +1,43 @@
+const test = require("node:test");
+const assert = require("node:assert/strict");
+
+const en = require("../../src/locales/en/translation.json");
+
+const load = () => import("../../src/components/notes/shared.ts");
+
+// Codes chunkedCloudTranscribe and interpretTranscribeResponse surface to the
+// upload UI. Anything added there needs a locale key or it reaches the user as
+// a raw English main-process string.
+const MAPPED_CODES = ["NO_SPEECH_DETECTED", "CHUNK_LOSS_EXCEEDED"];
+
+test("every mapped code resolves to a key that exists in en", async () => {
+ const { transcriptionErrorKey } = await load();
+
+ for (const code of MAPPED_CODES) {
+ const key = transcriptionErrorKey({ code });
+ assert.ok(key, `${code} has no i18n key`);
+ assert.ok(en.notes.upload[key], `notes.upload.${key} is missing from en`);
+ }
+});
+
+// OpenWhispr Cloud rethrows coded failures through withSessionRefresh, so the
+// upload views only see them in a catch block — resolving from the returned
+// result alone silently left cloud users with the raw English message.
+test("a thrown failure resolves the same key as a returned one", async () => {
+ const { transcriptionErrorKey } = await load();
+
+ const thrown = Object.assign(new Error("12 of 19 audio segments were lost"), {
+ code: "CHUNK_LOSS_EXCEEDED",
+ });
+ assert.equal(transcriptionErrorKey(thrown), "chunkLossExceeded");
+ assert.equal(transcriptionErrorKey({ code: "CHUNK_LOSS_EXCEEDED" }), "chunkLossExceeded");
+});
+
+test("unmapped, codeless and absent failures fall through to the raw message", async () => {
+ const { transcriptionErrorKey } = await load();
+
+ assert.equal(transcriptionErrorKey({ code: "SERVER_ERROR" }), undefined);
+ assert.equal(transcriptionErrorKey(new Error("boom")), undefined);
+ assert.equal(transcriptionErrorKey(null), undefined);
+ assert.equal(transcriptionErrorKey(undefined), undefined);
+});