Merge pull request #61 from DanielSWolf/bugfix/misc

Multiple small fixes and improvements

CHANGELOG.md

@@ -3,9 +3,11 @@
 ## Unreleased
 
 * **Added** basic support for non-English recordings through phonetic recognition ([issue #45](https://github.com/DanielSWolf/rhubarb-lip-sync/issues/45)).
+* **Improved** processing speed for WAVE files ([issue #58](https://github.com/DanielSWolf/rhubarb-lip-sync/issues/58)).
 * **Fixed** a bug that resulted in unwanted mouth movement at beginning of a recording ([issue #53](https://github.com/DanielSWolf/rhubarb-lip-sync/issues/53)).
 * **Fixed** a bug that garbled special characters in the output file path ([issue #54](https://github.com/DanielSWolf/rhubarb-lip-sync/issues/54)).
 * **Fixed** a bug that prevented the progress bar from reaching 100% ([issue #48](https://github.com/DanielSWolf/rhubarb-lip-sync/issues/48)).
+* **Fixed** file paths in exported XML and JSON files ([issue #59](https://github.com/DanielSWolf/rhubarb-lip-sync/issues/59)).
 
 ## Version 1.8.0
 

rhubarb/src/audio/WaveFileReader.cpp

@@ -192,7 +192,9 @@ SampleReader WaveFileReader::createUnsafeSampleReader() const {
 		](size_type index) mutable {
 		const std::streampos newFilePos = formatInfo.dataOffset
 			+ static_cast<std::streamoff>(index * formatInfo.bytesPerFrame);
+		if (newFilePos != filePos) {
 			file->seekg(newFilePos);
+		}
 		const value_type result =
 			readSample(*file, formatInfo.sampleFormat, formatInfo.channelCount);
 		filePos = newFilePos + static_cast<std::streamoff>(formatInfo.bytesPerFrame);

rhubarb/src/audio/voiceActivityDetection.cpp

@@ -75,6 +75,14 @@ JoiningBoundedTimeline<void> detectVoiceActivity(
 		}
 	}
 
+	// Discard very short segments of activity
+	const centiseconds minSegmentLength(5);
+	for (const auto& segment : activity) {
+		if (segment.getDuration() < minSegmentLength) {
+			activity.clear(segment.getTimeRange());
+		}
+	}
+
 	logging::debugFormat(
 		"Found {} sections of voice activity: {}",
 		activity.size(),

rhubarb/src/exporters/JsonExporter.cpp

@@ -1,6 +1,7 @@
 #include "JsonExporter.h"
 #include "exporterTools.h"
 #include "tools/stringTools.h"
+#include <boost/filesystem.hpp>
 
 using std::string;
 
@@ -10,7 +11,7 @@ void JsonExporter::exportAnimation(const ExporterInput& input, std::ostream& out
 	// the formatting.
 	outputStream << "{\n";
 	outputStream << "  \"metadata\": {\n";
-	outputStream << "    \"soundFile\": \"" << escapeJsonString(input.inputFilePath.string()) << "\",\n";
+	outputStream << "    \"soundFile\": \"" << escapeJsonString(absolute(input.inputFilePath).string()) << "\",\n";
 	outputStream << "    \"duration\": " << formatDuration(input.animation.getRange().getDuration()) << "\n";
 	outputStream << "  },\n";
 	outputStream << "  \"mouthCues\": [\n";

rhubarb/src/exporters/XmlExporter.cpp

@@ -3,6 +3,7 @@
 #include <boost/property_tree/xml_parser.hpp>
 #include <boost/version.hpp>
 #include "exporterTools.h"
+#include <boost/filesystem.hpp>
 
 using std::string;
 using boost::property_tree::ptree;
@@ -11,7 +12,7 @@ void XmlExporter::exportAnimation(const ExporterInput& input, std::ostream& outp
 	ptree tree;
 
 	// Add metadata
-	tree.put("rhubarbResult.metadata.soundFile", input.inputFilePath.string());
+	tree.put("rhubarbResult.metadata.soundFile", absolute(input.inputFilePath).string());
 	tree.put(
 		"rhubarbResult.metadata.duration",
 		formatDuration(input.animation.getRange().getDuration())

rhubarb/src/recognition/PhoneticRecognizer.cpp

@@ -25,6 +25,13 @@ static lambda_unique_ptr<ps_decoder_t> createDecoder(optional<std::string> dialo
 			// Low values (<= 0.4) can lead to fluttering animation.
 			// High values (>= 1.0) can lead to imprecise or freezing animation.
 			"-lw", "0.8",
+			// Add noise against zero silence
+			// (see http://cmusphinx.sourceforge.net/wiki/faq#qwhy_my_accuracy_is_poor)
+			"-dither", "yes",
+			// Disable VAD -- we're doing that ourselves
+			"-remove_silence", "no",
+			// Perform per-utterance cepstral mean normalization
+			"-cmn", "batch",
 
 			// The following settings are recommended at
 			// http://cmusphinx.sourceforge.net/wiki/phonemerecognition

rhubarb/src/recognition/pocketSphinxTools.cpp

@@ -219,10 +219,19 @@ BoundedTimeline<string> recognizeWords(const vector<int16_t>& audioBuffer, ps_de
 	BoundedTimeline<string> result(
 		TimeRange(0_cs, centiseconds(100 * audioBuffer.size() / sphinxSampleRate))
 	);
+	const bool phonetic = cmd_ln_boolean_r(decoder.config, "-allphone_ci");
+	if (!phonetic) {
+		// If the decoder is in word mode (as opposed to phonetic recognition), it expects each
+		// utterance to contain speech. If it doesn't, ps_seg_word() logs the annoying error
+		// "Couldn't find <s> in first frame".
+		// Not every utterance does contain speech, however. In this case, we exit early to prevent
+		// the log output.
+		// We *don't* to that in phonetic mode because here, the same code would omit valid phones.
 		const bool noWordsRecognized = reinterpret_cast<ngram_search_t*>(decoder.search)->bpidx == 0;
 		if (noWordsRecognized) {
 			return result;
 		}
+	}
 
 	// Collect words
 	for (ps_seg_t* it = ps_seg_iter(&decoder); it; it = ps_seg_next(it)) {