fix: recover empty short Whisper transcriptions

Author: Mr-Sunglasses

Sources/VocaMac/Services/WhisperService.swift

@@ -206,6 +206,22 @@ final class WhisperService: @unchecked Sendable {
                 fullText = Self.filterHallucinationTokens(rawText)
             }
 
+            // Whisper Tiny can return no tokens for valid sub-second speech.
+            // Retry only an empty short result with trailing silence so normal
+            // successful dictation keeps the single-pass fast path.
+            if let paddedAudio = Self.paddedAudioForShortEmptyTranscription(
+                audioData,
+                transcription: fullText
+            ) {
+                VocaLogger.warning(
+                    .whisperService,
+                    "Short transcription was empty for \(loadedModelName ?? "unknown model"); retrying with trailing silence"
+                )
+                results = try await kit.transcribe(audioArray: paddedAudio, decodeOptions: options)
+                rawText = results.map { $0.text }.joined(separator: " ")
+                fullText = Self.filterHallucinationTokens(rawText)
+            }
+
             let elapsed = CFAbsoluteTimeGetCurrent() - startTime
 
             // Get detected language from first result
@@ -293,6 +309,24 @@ final class WhisperService: @unchecked Sendable {
         promptTokens != nil && rawText.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
     }
 
+    /// Some Whisper models return no tokens for otherwise valid sub-second
+    /// speech. A small trailing-silence pad gives the retry more context without
+    /// changing the reported audio length.
+    static func paddedAudioForShortEmptyTranscription(
+        _ audio: [Float],
+        transcription: String
+    ) -> [Float]? {
+        let minimumSampleCount = 17_600 // 1.1 seconds at 16 kHz
+        guard transcription.isEmpty,
+              !audio.isEmpty,
+              audio.count < minimumSampleCount else {
+            return nil
+        }
+        var padded = audio
+        padded.append(contentsOf: repeatElement(0, count: minimumSampleCount - audio.count))
+        return padded
+    }
+
     /// Encode custom vocabulary into WhisperKit conditioning tokens.
     /// Returns nil when there are no terms or the tokenizer isn't ready yet.
     /// Framed as a "Glossary:" prompt, which nudges Whisper to treat the terms

Tests/VocaMacTests/WhisperServiceTests.swift

@@ -96,3 +96,33 @@ final class WhisperServiceVocabularyTests: XCTestCase {
         XCTAssertFalse(WhisperService.shouldRetryWithoutVocabulary(rawText: "", promptTokens: nil))
     }
 }
+
+// MARK: - WhisperService Short Audio Tests
+
+final class WhisperServiceShortAudioTests: XCTestCase {
+    func testEmptyShortTranscriptionPadsToMinimumDecoderWindow() throws {
+        let audio: [Float] = [0.25, -0.5, 0.75]
+        let padded = try XCTUnwrap(
+            WhisperService.paddedAudioForShortEmptyTranscription(audio, transcription: "")
+        )
+
+        XCTAssertEqual(padded.count, 17_600)
+        XCTAssertEqual(Array(padded.prefix(audio.count)), audio)
+        XCTAssertTrue(padded.dropFirst(audio.count).allSatisfy { $0 == 0 })
+    }
+
+    func testSuccessfulOrLongTranscriptionDoesNotRetry() {
+        XCTAssertNil(
+            WhisperService.paddedAudioForShortEmptyTranscription([0.5], transcription: "Hello")
+        )
+        XCTAssertNil(
+            WhisperService.paddedAudioForShortEmptyTranscription(
+                Array(repeating: 0.5, count: 17_600),
+                transcription: ""
+            )
+        )
+        XCTAssertNil(
+            WhisperService.paddedAudioForShortEmptyTranscription([], transcription: "")
+        )
+    }
+}