server: return 400 for an inline voice_ref that is not a WAV

app/cli/request.cpp

@@ -380,7 +380,11 @@ engine::runtime::AudioBuffer read_voice_ref(
     if (bytes.size() > kMaxVoiceRefBytes) {
         throw InvalidRequestError("voice_ref base64 data exceeds the 5 MiB limit");
     }
-    return read_audio_buffer(std::string_view(reinterpret_cast<const char *>(bytes.data()), bytes.size()));
+    try {
+        return read_audio_buffer(std::string_view(reinterpret_cast<const char *>(bytes.data()), bytes.size()));
+    } catch (const std::runtime_error & error) {
+        throw InvalidRequestError(std::string("voice_ref base64 data is not a readable WAV (") + error.what() + ")");
+    }
 }
 
 }  // namespace

app/server/api.md

@@ -455,7 +455,7 @@ path, `{"type": "path", "path": "..."}`, or `{"type": "base64", "data": "..."}`
 with a Base64-encoded WAV or `data:audio/wav;base64,...` URI of at most 5 MiB.
 The inline form lets a client send a reference voice, for example the target
 speaker of a voice conversion, without the server having to store it on disk.
-An invalid `voice_ref` returns HTTP 400. CLI request files and the batch and
+A malformed `voice_ref`, or inline data that is not a readable WAV, returns HTTP 400. CLI request files and the batch and
 stream endpoints share this request format, so they accept it too.
 
 A result's artifacts (opaque payloads such as speaker embeddings or acoustic

tests/unittests/test_cli_request_options.cpp

@@ -580,6 +580,11 @@ void test_request_json_voice_ref() {
     require_voice_ref_rejected("{\"type\":\"base64\"}", "needs a data string");
     require_voice_ref_rejected("{\"type\":\"base64\",\"data\":\"@@@@\"}", "not valid base64");
     require_voice_ref_rejected("{\"type\":\"base64\",\"data\":\"\"}", "empty payload");
+    const std::string not_wav = "plain text, not a WAV file";
+    require_voice_ref_rejected(
+        "{\"type\":\"base64\",\"data\":\"" +
+            minitts::app::base64_encode(std::vector<uint8_t>(not_wav.begin(), not_wav.end())) + "\"}",
+        "not a readable WAV");
     require_voice_ref_rejected(
         "{\"type\":\"base64\",\"data\":\"" + std::string(size_t{8} * 1024 * 1024, 'A') + "\"}", "5 MiB limit");
 }