Previous bug (pre-fix): both providers hardcoded {@code "audio.ogg"} - * as the default filename even when the upstream content was WebM/Opus, - * which DashScope's HTTP path then tried to decode as Ogg and 400'd. - * That bug + DashScope's HTTP STT are both gone now (DashScope went to - * WebSocket); this class survives because Whisper still cares. + *
Previous bug (pre-fix): providers hardcoded {@code "audio.ogg"} as the + * default filename even when upstream content was WebM/Opus. The helper now + * serves both multipart filenames (Whisper-compatible endpoints) and MIME- + * qualified data URLs (Qwen3-ASR). */ public final class AudioMimeTypes { @@ -70,6 +69,26 @@ public final class AudioMimeTypes { return "audio." + (extension != null ? extension : DEFAULT_EXTENSION); } + /** + * Resolve the MIME type that describes the actual encoded bytes. + * + *
Data-URL based APIs (notably Qwen3-ASR) inspect the media type in + * {@code data:audio/wav;base64,...}. Sending an empty media type can make + * the service mis-detect or truncate otherwise valid audio while still + * returning HTTP 200, so callers must never emit {@code data:;base64,...}. + */ + public static String resolveContentType(String fileName, String contentType) { + String extension = extensionForContentType(contentType); + if (extension != null) { + return EXTENSION_TO_CONTENT_TYPE.get(extension); + } + String resolvedFileName = resolveFileName(fileName, null); + String fileExtension = extensionOf(resolvedFileName); + return fileExtension != null + ? EXTENSION_TO_CONTENT_TYPE.get(fileExtension) + : EXTENSION_TO_CONTENT_TYPE.get(DEFAULT_EXTENSION); + } + /** Extract the lower-cased extension (without the dot), or null. Package-private for tests. */ static String extensionOf(String fileName) { if (fileName == null) return null; diff --git a/mateclaw-server/src/main/java/vip/mate/stt/WavPcmExtractor.java b/mateclaw-server/src/main/java/vip/mate/stt/WavPcmExtractor.java index e357477a..efec3c07 100644 --- a/mateclaw-server/src/main/java/vip/mate/stt/WavPcmExtractor.java +++ b/mateclaw-server/src/main/java/vip/mate/stt/WavPcmExtractor.java @@ -30,14 +30,20 @@ public final class WavPcmExtractor { private WavPcmExtractor() {} /** - * True when the bytes carry the RIFF/WAVE magic and are long enough to - * hold the canonical 44-byte header. Cheap gate for callers that only - * want PCM diagnostics on inputs {@link #extract} can actually handle. + * True only for the 44-byte PCM16/mono layout produced by MateClaw's web + * recorder. Stereo WAVs and files with extra chunks are still valid audio, + * but callers must send them directly to STT instead of applying the + * mono-specific sample math in this helper. */ public static boolean isCanonicalWav(byte[] bytes) { return bytes != null && bytes.length >= CANONICAL_HEADER_BYTES && bytes[0] == 'R' && bytes[1] == 'I' && bytes[2] == 'F' && bytes[3] == 'F' - && bytes[8] == 'W' && bytes[9] == 'A' && bytes[10] == 'V' && bytes[11] == 'E'; + && bytes[8] == 'W' && bytes[9] == 'A' && bytes[10] == 'V' && bytes[11] == 'E' + && bytes[12] == 'f' && bytes[13] == 'm' && bytes[14] == 't' && bytes[15] == ' ' + && unsignedShort(bytes, 20) == 1 + && unsignedShort(bytes, 22) == 1 + && unsignedShort(bytes, 34) == 16 + && bytes[36] == 'd' && bytes[37] == 'a' && bytes[38] == 't' && bytes[39] == 'a'; } /** @@ -74,4 +80,8 @@ public final class WavPcmExtractor { .getInt(); } + private static int unsignedShort(byte[] bytes, int offset) { + return (bytes[offset] & 0xFF) | ((bytes[offset + 1] & 0xFF) << 8); + } + } diff --git a/mateclaw-server/src/main/java/vip/mate/stt/provider/DashScopeSttProvider.java b/mateclaw-server/src/main/java/vip/mate/stt/provider/DashScopeSttProvider.java index 68985c54..bbec76a8 100644 --- a/mateclaw-server/src/main/java/vip/mate/stt/provider/DashScopeSttProvider.java +++ b/mateclaw-server/src/main/java/vip/mate/stt/provider/DashScopeSttProvider.java @@ -57,7 +57,7 @@ import java.util.Map; *
{@code
* {"model":"qwen3-asr-flash",
* "messages":[{"role":"user","content":[
- * {"type":"input_audio","input_audio":{"data":"data:;base64,...","format":"wav"}}]}],
+ * {"type":"input_audio","input_audio":{"data":"data:audio/wav;base64,..."}}]}],
* "stream":false,
* "asr_options":{"language":"zh"}} // omitted → auto language detection
* }
@@ -80,6 +80,12 @@ public class DashScopeSttProvider implements SttProvider {
/** Overall budget for the single HTTP round trip. */
static final int HTTP_TIMEOUT_MS = 60_000;
+ /** Qwen3-ASR-Flash OpenAI-compatible request limit. */
+ static final int MAX_AUDIO_BYTES = 10 * 1024 * 1024;
+
+ /** Reject electrical noise that would otherwise be hallucinated as a filler such as “嗯”. */
+ static final int MIN_SPEECH_RMS = 16;
+
private final ModelProviderService modelProviderService;
private final ObjectMapper objectMapper;
@@ -122,29 +128,40 @@ public class DashScopeSttProvider implements SttProvider {
if (audio == null || audio.length == 0) {
return SttResult.failure("音频为空");
}
+ if (audio.length > MAX_AUDIO_BYTES) {
+ return SttResult.failure("音频超过 Qwen3-ASR 10 MB 限制");
+ }
// Silence gate — only for WAV, where we can read PCM directly.
// "Mic captured nothing" is by far the most common voice-input
// failure; catching it here yields a precise error instead of an
// empty transcript from the model. Non-WAV inputs (IM voice
// notes) skip the gate and go straight to the API.
+ double localDurationSeconds = -1;
if (WavPcmExtractor.isCanonicalWav(audio)) {
byte[] pcm = WavPcmExtractor.extract(audio);
int[] peakRms = computePcmPeakRms(pcm);
- if (peakRms[0] == 0) {
- log.warn("[DashScope STT] PCM is silent (peak=0, bytes={}) — check mic permission / frontend recording",
- pcm.length);
- return SttResult.failure(
- "音频为静音（PCM peak=0）— 检查麦克风权限或前端录制实现");
+ int sampleRate = WavPcmExtractor.sampleRate(audio);
+ if (sampleRate <= 0) {
+ return SttResult.failure("WAV 采样率无效: " + sampleRate);
}
- log.debug("[DashScope STT] PCM stats — bytes={} peak={} rms={}",
- pcm.length, peakRms[0], peakRms[1]);
+ localDurationSeconds = (double) pcm.length / (sampleRate * 2L);
+ if (peakRms[1] < MIN_SPEECH_RMS) {
+ log.warn("[DashScope STT] PCM is silent/near-silent (peak={}, rms={}, bytes={}) — check mic permission / frontend recording",
+ peakRms[0], peakRms[1], pcm.length);
+ return SttResult.failure(
+ "音频为静音或音量过低（PCM peak=" + peakRms[0]
+ + ", rms=" + peakRms[1] + "）— 请检查麦克风权限和输入音量");
+ }
+ log.debug("[DashScope STT] PCM stats — bytes={} peak={} rms={} duration={}s",
+ pcm.length, peakRms[0], peakRms[1], localDurationSeconds);
}
String model = (request.getModel() != null && !request.getModel().isBlank())
? request.getModel() : DEFAULT_MODEL;
- String format = resolveFormat(request.getFileName(), request.getContentType());
- String body = buildRequestBody(model, audio, format, request.getLanguage());
+ String mimeType = AudioMimeTypes.resolveContentType(
+ request.getFileName(), request.getContentType());
+ String body = buildRequestBody(model, audio, mimeType, request.getLanguage());
HttpResponse response = HttpRequest.post(ASR_ENDPOINT)
.header("Authorization", "Bearer " + apiKey.trim())
@@ -180,8 +197,19 @@ public class DashScopeSttProvider implements SttProvider {
}
String text = parseTranscript(responseBody);
- log.info("[DashScope STT] Transcribed {} chars (model={}, format={}, audioBytes={})",
- text.length(), model, format, audio.length);
+ if (text.isBlank()) {
+ return SttResult.failure("DashScope 未返回识别文本，请检查录音内容和输入音量");
+ }
+ int recognizedSeconds = parseRecognizedSeconds(responseBody);
+ if (isSuspiciouslyTruncated(localDurationSeconds, recognizedSeconds)) {
+ log.warn("[DashScope STT] decoded duration mismatch — local={}s remote={}s, mimeType={}, bytes={}",
+ localDurationSeconds, recognizedSeconds, mimeType, audio.length);
+ return SttResult.failure("DashScope 仅解码了约 " + recognizedSeconds
+ + " 秒音频，但本地录音约 " + Math.round(localDurationSeconds)
+ + " 秒；请检查录音编码或网关是否截断了音频");
+ }
+ log.info("[DashScope STT] Transcribed {} chars (model={}, mimeType={}, audioBytes={}, localDuration={}s, recognizedDuration={}s)",
+ text.length(), model, mimeType, audio.length, localDurationSeconds, recognizedSeconds);
return SttResult.success(text);
} catch (Exception e) {
log.error("[DashScope STT] Error: {}", e.getMessage(), e);
@@ -195,13 +223,14 @@ public class DashScopeSttProvider implements SttProvider {
/**
* Build the recognition request. The audio rides in a
- * {@code data:;base64,} URI — the separate {@code format} field tells the
- * service how to decode it, so the URI needs no media type.
+ * MIME-qualified data URI. Qwen3-ASR uses the media type to decode the
+ * file; unlike Qwen audio/translation models it does not define a
+ * separate {@code input_audio.format} request field.
*/
- String buildRequestBody(String model, byte[] audio, String format, String language) throws Exception {
+ String buildRequestBody(String model, byte[] audio, String mimeType, String language) throws Exception {
MapWhy this exists: MediaRecorder produces WebM/Opus, which DashScope - * Paraformer rejects (it accepts wav/mp3/m4a/flac/aac/amr/ogg-Vorbis only). - * OpenAI Whisper claims WebM support but is finicky with the codecs string - * MediaRecorder picks. WAV is the lowest common denominator that every STT - * provider accepts without server-side transcoding. + *
Why this exists: MediaRecorder output and codec strings vary by browser. + * A canonical PCM WAV is the lowest common denominator that every STT + * provider accepts without server-side transcoding, and its samples can be + * inspected locally for silence and duration before an API call. * *
Trade-off: WAV files are ~10x larger than Opus. For typical
* conversational STT (5-30s clips at 16 kHz / 16-bit / mono) that's
@@ -54,6 +53,12 @@ export interface WavRecording {
export class WavRecorder {
private audioContext: AudioContext | null = null;
private mediaStream: MediaStream | null = null;
+ /** Deduplicates warm-up/start getUserMedia calls so they cannot overwrite each other's stream. */
+ private mediaStreamPromise: Promise Safe to call multiple times. Subsequent calls return immediately.
*/
async warmUp(): Promise