feat(voice): modalità dialogo — voce continua hands-free (v0.2.0)

- Nuovo VoiceRecorder (QAudioInput, PCM 16 bit mono): analisi del livello in tempo reale e stop automatico dopo il silenzio (pausa e soglia configurabili dalle impostazioni); produce WAV per /api/transcribe; emette failed("no-speech") se non rileva voce (cattura scartata) - ChatPage: modalità dialogo (menu "Voice mode"): ascolto -> trascrizione -> invio -> lettura della risposta -> riascolto in ciclo; tocco sul testo di stato per uscire; limite di 3 errori consecutivi poi esce - Dettatura manuale: invio automatico dopo la trascrizione (opzione, default attivo, disattivabile) - Settings: sezione "Voice dialog" (auto-invio, pausa auto-stop, sensibilità microfono) - Il percorso di dettatura manuale (Recorder OGG) resta invariato
2026-09-12 09:14:25 +02:00
parent 479edb0f99
commit 9deeedf5b3
12 changed files with 620 additions and 18 deletions
@@ -29,6 +29,6 @@ speaks (validated live against a real instance).

 ## Status

-Version 0.1.2 — protocol and C++ core validated against a real `hermes-webui`
+Version 0.2.0 — protocol and C++ core validated against a real `hermes-webui`
 (login, sessions, streaming chat, TTS, STT); on-device testing (audio paths,
 Sailjail prompts, keyboard handling) still pending.
@@ -76,12 +76,12 @@ richiede il probe WebUI su :8899). Esito 12/09/2026: unit 9/9, live completo OK.

 ```bash
 cd /home/kaneda/workspace
-tar czf harbour-hermes-0.1.2.tar.gz \
-    --transform 's,^harbour-hermes,harbour-hermes-0.1.2,' \
+tar czf harbour-hermes-0.2.0.tar.gz \
+    --transform 's,^harbour-hermes,harbour-hermes-0.2.0,' \
    --exclude='.git' harbour-hermes
 ```

-Lo spec si aspetta la directory `harbour-hermes-0.1.2/` (pattern degli altri
+Lo spec si aspetta la directory `harbour-hermes-0.2.0/` (pattern degli altri
 progetti: cercato in modo robusto anche per i sorgenti live di sfdk).

 ## 5. Dipendenze lato server (WebUI)
@@ -16,6 +16,7 @@ SOURCES += \
    src/chatmodel.cpp \
    src/sessionsmodel.cpp \
    src/recorder.cpp \
+    src/voicerecorder.cpp \
    src/player.cpp

 HEADERS += \
@@ -26,6 +27,7 @@ HEADERS += \
    src/chatmodel.h \
    src/sessionsmodel.h \
    src/recorder.h \
+    src/voicerecorder.h \
    src/player.h

 OTHER_FILES += \
@@ -7,6 +7,11 @@ Page {

    property string sessionId
    property string noticeText: ""
+    // Modalità dialogo (voce continua): ascolto → trascrizione → invio →
+    // lettura della risposta → riascolto (vedi toggleVoiceMode()).
+    property bool voiceMode: false
+    property string voiceState: "idle"   // listening|transcribing|sending|streaming|speaking
+    property int voiceErrors: 0
    // Altezza tastiera: alza il composer quando la tastiera e' aperta.
    property real kbHeight: Qt.inputMethod.visible
        ? Math.max(0, page.height - Qt.inputMethod.keyboardRectangle.y)
@@ -49,6 +54,10 @@ Page {
        }

        PullDownMenu {
+            MenuItem {
+                text: page.voiceMode ? qsTr("Voice mode: on") : qsTr("Voice mode: off")
+                onClicked: page.toggleVoiceMode()
+            }
            MenuItem {
                text: qsTr("Read last reply aloud")
                onClicked: page.readLast()
@@ -119,18 +128,28 @@ Page {
            spacing: Theme.paddingSmall / 2

            Label {
+                id: statusLbl
                x: Theme.horizontalPageMargin
                width: parent.width - 2 * Theme.horizontalPageMargin
                font.pixelSize: Theme.fontSizeExtraSmall
-                color: recorder.recording ? Theme.highlightColor
-                                          : Theme.secondaryHighlightColor
-                visible: recorder.recording || api.transcribing || page.noticeText.length > 0
+                color: page.voiceMode ? Theme.highlightColor
+                                      : (recorder.recording ? Theme.highlightColor
+                                                            : Theme.secondaryHighlightColor)
+                visible: page.voiceMode || recorder.recording || api.transcribing || page.noticeText.length > 0
                truncationMode: TruncationMode.Fade
-                text: recorder.recording
-                      ? qsTr("Recording… %1 — tap Stop to transcribe").arg(page.formatMs(recorder.durationMs))
-                      : api.transcribing
-                        ? qsTr("Transcribing…")
-                        : page.noticeText
+                text: page.voiceMode
+                      ? page.voiceStatusText()
+                      : (recorder.recording
+                         ? qsTr("Recording… %1 — tap Stop to transcribe").arg(page.formatMs(recorder.durationMs))
+                         : (api.transcribing ? qsTr("Transcribing…") : page.noticeText))
+
+                // In modalità dialogo: un tocco sul testo esce (MouseArea
+                // interna al Label: nessuna interferenza con la Column).
+                MouseArea {
+                    anchors.fill: parent
+                    enabled: page.voiceMode
+                    onClicked: page.exitVoiceMode()
+                }
            }

            Row {
@@ -142,7 +161,7 @@ Page {
                    id: micBtn
                    text: recorder.recording ? qsTr("Stop") : qsTr("Mic")
                    // Button non ha "highlighted": lo stato è nel testo (Mic/Stop)
-                    enabled: !api.transcribing && !chat.busy || recorder.recording
+                    enabled: !page.voiceMode && (!api.transcribing && !chat.busy || recorder.recording)
                    onClicked: {
                        if (recorder.recording)
                            recorder.stop()
@@ -184,11 +203,30 @@ Page {
        onTriggered: page.noticeText = ""
    }

+    Timer {
+        id: voiceRetry
+        interval: 1200
+        onTriggered: {
+            if (page.voiceMode)
+                page.startListening()
+        }
+    }
+
    Connections {
        target: chat
        onRowCountChanged: scrollDebounce.restart()
        onStreamTick: scrollDebounce.restart()
        onAssistantFinished: {
+            if (page.voiceMode) {
+                var t = chat.lastAssistantText()
+                if (t.length > 0) {
+                    page.voiceState = "speaking"
+                    api.speak(t)
+                } else {
+                    voiceRetry.restart()
+                }
+                return
+            }
            if (appSettings.ttsAutoRead)
                page.readLast()
        }
@@ -198,11 +236,39 @@ Page {
    Connections {
        target: api
        onTranscriptionReady: {
+            page.voiceErrors = 0
+            if (page.voiceMode) {
+                if (text.trim().length === 0) {
+                    voiceRetry.restart()
+                    return
+                }
+                input.text = text
+                page.voiceState = "sending"
+                page.doSend()
+                return
+            }
            input.text = text
            input.focus = true
+            if (appSettings.dictationAutoSend)
+                page.doSend()
+        }
+        onTranscriptionFailed: {
+            page.showNotice(error)
+            if (page.voiceMode) {
+                page.voiceErrors += 1
+                if (page.voiceErrors >= 3) {
+                    page.exitVoiceMode()
+                } else {
+                    voiceRetry.restart()
+                }
+            }
+        }
+        onTtsFailed: {
+            if (page.voiceMode)
+                voiceRetry.restart()
+            else
+                page.showNotice(error)
        }
-        onTranscriptionFailed: page.showNotice(error)
-        onTtsFailed: page.showNotice(error)
    }

    Component.onCompleted: {
@@ -210,6 +276,102 @@ Page {
            chat.openSession(sessionId)
    }

+    Component.onDestruction: {
+        if (page.voiceMode) {
+            page.voiceMode = false
+            voiceRec.stop()
+            player.stop()
+        }
+    }
+
+    Connections {
+        target: player
+        onFinished: {
+            if (page.voiceMode)
+                page.startListening()
+        }
+    }
+
+    Connections {
+        target: voiceRec
+        onListening: {
+            if (page.voiceMode)
+                page.voiceState = "listening"
+        }
+        onFinished: {
+            if (!page.voiceMode)
+                return
+            if (wavPath.length === 0)
+                return
+            page.voiceState = "transcribing"
+            api.transcribeFile(wavPath)
+        }
+        onFailed: {
+            if (error === "no-speech") {
+                if (page.voiceMode)
+                    voiceRetry.restart()
+                return
+            }
+            page.showNotice(error)
+            if (page.voiceMode) {
+                page.voiceErrors += 1
+                if (page.voiceErrors >= 3) {
+                    page.exitVoiceMode()
+                    page.showNotice(qsTr("Voice mode off after repeated errors"))
+                } else {
+                    voiceRetry.restart()
+                }
+            }
+        }
+    }
+
+    function toggleVoiceMode() {
+        if (page.voiceMode)
+            page.exitVoiceMode()
+        else
+            page.enterVoiceMode()
+    }
+
+    function enterVoiceMode() {
+        if (chat.busy || recorder.recording) {
+            page.showNotice(qsTr("Wait for the current reply first"))
+            return
+        }
+        player.stop()
+        page.voiceErrors = 0
+        page.voiceMode = true
+        page.voiceState = "listening"
+        voiceRec.start()
+    }
+
+    function exitVoiceMode() {
+        page.voiceMode = false
+        page.voiceState = "idle"
+        voiceRec.stop()
+        player.stop()
+    }
+
+    function startListening() {
+        if (!page.voiceMode)
+            return
+        page.voiceState = "listening"
+        voiceRec.start()
+    }
+
+    function voiceStatusText() {
+        if (page.voiceState === "listening")
+            return qsTr("● Listening — speak, I stop at the pause (tap to exit)")
+        if (page.voiceState === "transcribing")
+            return qsTr("Transcribing…")
+        if (page.voiceState === "sending")
+            return qsTr("Sending…")
+        if (page.voiceState === "streaming")
+            return qsTr("Hermes is writing…")
+        if (page.voiceState === "speaking")
+            return qsTr("Speaking…")
+        return qsTr("Voice mode on")
+    }
+
    function doSend() {
        var t = input.text
        if (t.trim().length === 0)
@@ -94,6 +94,40 @@ Page {
                onCheckedChanged: appSettings.ttsAutoRead = checked
            }

+            SectionHeader { text: qsTr("Voice dialog") }
+
+            TextSwitch {
+                text: qsTr("Send right after dictation")
+                description: qsTr("Send the transcribed text automatically, without tapping Send")
+                checked: appSettings.dictationAutoSend
+                onCheckedChanged: appSettings.dictationAutoSend = checked
+            }
+
+            ComboBox {
+                label: qsTr("Auto-stop pause")
+                currentIndex: {
+                    var v = appSettings.vadSilenceMs
+                    return v <= 1000 ? 0 : (v <= 1500 ? 1 : (v <= 2000 ? 2 : 3))
+                }
+                menu: ContextMenu {
+                    MenuItem { text: qsTr("1.0 second"); onClicked: appSettings.vadSilenceMs = 1000 }
+                    MenuItem { text: qsTr("1.5 seconds"); onClicked: appSettings.vadSilenceMs = 1500 }
+                    MenuItem { text: qsTr("2.0 seconds"); onClicked: appSettings.vadSilenceMs = 2000 }
+                    MenuItem { text: qsTr("3.0 seconds"); onClicked: appSettings.vadSilenceMs = 3000 }
+                }
+            }
+
+            ComboBox {
+                label: qsTr("Mic sensitivity")
+                currentIndex: appSettings.vadThreshold <= 500 ? 0
+                              : (appSettings.vadThreshold <= 1000 ? 1 : 2)
+                menu: ContextMenu {
+                    MenuItem { text: qsTr("High (even soft voices)"); onClicked: appSettings.vadThreshold = 400 }
+                    MenuItem { text: qsTr("Medium"); onClicked: appSettings.vadThreshold = 800 }
+                    MenuItem { text: qsTr("Low (noisy rooms)"); onClicked: appSettings.vadThreshold = 1600 }
+                }
+            }
+
            SectionHeader { text: qsTr("Account") }

            Button {
@@ -1,6 +1,6 @@
 Name:       harbour-hermes
 Summary:    Client for the Hermes Web UI with voice
-Version:    0.1.2
+Version:    0.2.0
 Release:    1
 Group:      Qt/Qt
 License:    MIT
@@ -24,6 +24,11 @@ Sessions list, chat with live streaming replies, voice dictation
 (server-side TTS), plus a hands-free conversation mode.

 %changelog
+* Sat Sep 12 2026 Carlo Baratto <carlo@carlobaratto.it> - 0.2.0-1
+- Modalita' dialogo (voce continua): ascolto con stop automatico sul
+  silenzio (QAudioInput + VAD), trascrizione, invio automatico, lettura
+  della risposta e riascolto in ciclo; pausa e sensibilita' microfono
+  configurabili; la dettatura manuale puo' inviare da sola
 * Sat Sep 12 2026 Carlo Baratto <carlo@carlobaratto.it> - 0.1.2-1
 - Fix lista sessioni vuota sul device: le QVariantList di mappe non
  espongono i campi nei delegate su Qt 5.6 -> nuovo SessionsModel
@@ -9,6 +9,7 @@
 #include "player.h"
 #include "recorder.h"
 #include "settings.h"
+#include "voicerecorder.h"

 Q_DECL_EXPORT int main(int argc, char *argv[])
 {
@@ -18,15 +19,16 @@ Q_DECL_EXPORT int main(int argc, char *argv[])
    // AppConfigLocation = ~/.config/harbour/hermes (zona persistente della sandbox).
    app->setOrganizationName(QStringLiteral("harbour"));
    app->setApplicationName(QStringLiteral("hermes"));
-    app->setApplicationVersion(QStringLiteral("0.1.0"));
+    app->setApplicationVersion(QStringLiteral("0.2.0"));

-    qDebug() << "harbour-hermes v0.1.0 build" << __DATE__ << __TIME__;
+    qDebug() << "harbour-hermes v0.2.0 build" << __DATE__ << __TIME__;

    Settings settings;
    ApiClient api(&settings);
    ChatModel chat(&api);
    Recorder recorder;
    Player player;
+    VoiceRecorder voiceRec(&settings);

    // Flusso voce: registrazione -> trascrizione; TTS pronto -> riproduzione.
    QObject::connect(&recorder, &Recorder::finished, &api, &ApiClient::transcribeFile);
@@ -38,6 +40,7 @@ Q_DECL_EXPORT int main(int argc, char *argv[])
    view->rootContext()->setContextProperty(QStringLiteral("chat"), &chat);
    view->rootContext()->setContextProperty(QStringLiteral("recorder"), &recorder);
    view->rootContext()->setContextProperty(QStringLiteral("player"), &player);
+    view->rootContext()->setContextProperty(QStringLiteral("voiceRec"), &voiceRec);
    view->rootContext()->setContextProperty(QStringLiteral("appVersion"),
                                            QCoreApplication::applicationVersion());

@@ -70,6 +70,45 @@ void Settings::setTtsAutoRead(bool on)
    emit ttsAutoReadChanged();
 }

+bool Settings::dictationAutoSend() const
+{
+    return m_settings.value(QStringLiteral("dictationAutoSend"), true).toBool();
+}
+
+void Settings::setDictationAutoSend(bool on)
+{
+    if (on == dictationAutoSend())
+        return;
+    m_settings.setValue(QStringLiteral("dictationAutoSend"), on);
+    emit dictationAutoSendChanged();
+}
+
+int Settings::vadSilenceMs() const
+{
+    return m_settings.value(QStringLiteral("vadSilenceMs"), 1500).toInt();
+}
+
+void Settings::setVadSilenceMs(int ms)
+{
+    if (ms == vadSilenceMs())
+        return;
+    m_settings.setValue(QStringLiteral("vadSilenceMs"), ms);
+    emit vadSilenceMsChanged();
+}
+
+int Settings::vadThreshold() const
+{
+    return m_settings.value(QStringLiteral("vadThreshold"), 800).toInt();
+}
+
+void Settings::setVadThreshold(int threshold)
+{
+    if (threshold == vadThreshold())
+        return;
+    m_settings.setValue(QStringLiteral("vadThreshold"), threshold);
+    emit vadThresholdChanged();
+}
+
 QString Settings::defaultWorkspace() const
 {
    return m_settings.value(QStringLiteral("defaultWorkspace"), QString()).toString();
@@ -19,6 +19,12 @@ class Settings : public QObject
    Q_PROPERTY(QString ttsVoice READ ttsVoice WRITE setTtsVoice NOTIFY ttsVoiceChanged)
    Q_PROPERTY(QString ttsEngine READ ttsEngine WRITE setTtsEngine NOTIFY ttsEngineChanged)
    Q_PROPERTY(bool ttsAutoRead READ ttsAutoRead WRITE setTtsAutoRead NOTIFY ttsAutoReadChanged)
+    Q_PROPERTY(bool dictationAutoSend READ dictationAutoSend WRITE setDictationAutoSend
+               NOTIFY dictationAutoSendChanged)
+    Q_PROPERTY(int vadSilenceMs READ vadSilenceMs WRITE setVadSilenceMs
+               NOTIFY vadSilenceMsChanged)
+    Q_PROPERTY(int vadThreshold READ vadThreshold WRITE setVadThreshold
+               NOTIFY vadThresholdChanged)
    Q_PROPERTY(QString defaultWorkspace READ defaultWorkspace WRITE setDefaultWorkspace
               NOTIFY defaultWorkspaceChanged)

@@ -37,6 +43,15 @@ public:
    bool ttsAutoRead() const;
    void setTtsAutoRead(bool on);

+    bool dictationAutoSend() const;
+    void setDictationAutoSend(bool on);
+
+    int vadSilenceMs() const;
+    void setVadSilenceMs(int ms);
+
+    int vadThreshold() const;
+    void setVadThreshold(int threshold);
+
    QString defaultWorkspace() const;
    void setDefaultWorkspace(const QString &ws);

@@ -45,6 +60,9 @@ signals:
    void ttsVoiceChanged();
    void ttsEngineChanged();
    void ttsAutoReadChanged();
+    void dictationAutoSendChanged();
+    void vadSilenceMsChanged();
+    void vadThresholdChanged();
    void defaultWorkspaceChanged();

 private:
@@ -0,0 +1,264 @@
+#include "voicerecorder.h"
+
+#include <QAudioDeviceInfo>
+#include <QAudioInput>
+#include <QDateTime>
+#include <QDebug>
+#include <QDir>
+#include <QFile>
+#include <QIODevice>
+#include <QStandardPaths>
+#include <QTimer>
+
+#include "settings.h"
+
+namespace {
+
+void append16(QByteArray &b, quint16 v)
+{
+    b.append(char(v & 0xff));
+    b.append(char((v >> 8) & 0xff));
+}
+
+void append32(QByteArray &b, quint32 v)
+{
+    b.append(char(v & 0xff));
+    b.append(char((v >> 8) & 0xff));
+    b.append(char((v >> 16) & 0xff));
+    b.append(char((v >> 24) & 0xff));
+}
+
+// Picco (valore assoluto massimo) dei campioni 16 bit little-endian del blocco.
+int peakOf(const QByteArray &chunk)
+{
+    int peak = 0;
+    const char *d = chunk.constData();
+    const int n = chunk.size() - (chunk.size() % 2);
+    for (int i = 0; i < n; i += 2) {
+        const qint16 s = qint16(quint8(d[i]) | (quint8(d[i + 1]) << 8));
+        const int v = qAbs(int(s));
+        if (v > peak)
+            peak = v;
+    }
+    return peak;
+}
+
+} // namespace
+
+VoiceRecorder::VoiceRecorder(Settings *settings, QObject *parent)
+    : QObject(parent)
+    , m_settings(settings)
+    , m_input(nullptr)
+    , m_device(nullptr)
+    , m_silenceTimer(new QTimer(this))
+    , m_lastVoiceMs(0)
+    , m_lastLevelEmitMs(0)
+    , m_hasVoice(false)
+    , m_active(false)
+    , m_level(0.0)
+{
+    m_silenceTimer->setInterval(150);
+    connect(m_silenceTimer, &QTimer::timeout, this, &VoiceRecorder::onSilenceCheck);
+}
+
+void VoiceRecorder::start()
+{
+    if (m_active)
+        return;
+
+    const QAudioDeviceInfo info = QAudioDeviceInfo::defaultInputDevice();
+    if (info.isNull()) {
+        emit failed(tr("No audio input device"));
+        return;
+    }
+
+    // PCM 16 bit mono: 16 kHz di solito basta per la voce; altrimenti si
+    // ripiega su frequenze piu' comuni mantenendo il 16 bit mono.
+    QAudioFormat format;
+    format.setChannelCount(1);
+    format.setSampleSize(16);
+    format.setCodec(QStringLiteral("audio/pcm"));
+    format.setByteOrder(QAudioFormat::LittleEndian);
+    format.setSampleType(QAudioFormat::SignedInt);
+
+    const int rates[] = { 16000, 48000, 44100, 32000, 8000 };
+    bool found = false;
+    for (unsigned i = 0; i < sizeof(rates) / sizeof(rates[0]); ++i) {
+        format.setSampleRate(rates[i]);
+        if (info.isFormatSupported(format)) {
+            found = true;
+            break;
+        }
+    }
+    if (!found) {
+        emit failed(tr("Audio input format not supported"));
+        return;
+    }
+
+    m_format = format;
+    m_pcm.clear();
+    m_hasVoice = false;
+    m_level = 0.0;
+    m_lastVoiceMs = 0;
+    m_lastLevelEmitMs = 0;
+
+    m_input = new QAudioInput(info, m_format, this);
+    m_device = m_input->start();
+    if (!m_device) {
+        delete m_input;
+        m_input = nullptr;
+        emit failed(tr("Could not start audio capture"));
+        return;
+    }
+    connect(m_device, &QIODevice::readyRead, this, &VoiceRecorder::onReadyRead);
+
+    m_elapsed.start();
+    m_active = true;
+    emit activeChanged();
+    emit levelChanged();
+    emit listening();
+    qDebug() << "[voice] ascolto avviato:" << m_format.sampleRate() << "Hz mono 16 bit";
+    m_silenceTimer->start();
+}
+
+void VoiceRecorder::stop()
+{
+    if (!m_active)
+        return;
+    finalize();
+}
+
+void VoiceRecorder::onReadyRead()
+{
+    if (!m_device)
+        return;
+    const QByteArray chunk = m_device->readAll();
+    if (chunk.isEmpty())
+        return;
+    m_pcm.append(chunk);
+
+    const int peak = peakOf(chunk);
+    const qint64 elapsed = m_elapsed.elapsed();
+    if (peak > m_settings->vadThreshold()) {
+        m_lastVoiceMs = elapsed;
+        m_hasVoice = true;
+    }
+
+    if (elapsed - m_lastLevelEmitMs >= 150) {
+        m_lastLevelEmitMs = elapsed;
+        m_level = qreal(peak) / 32768.0;
+        emit levelChanged();
+    }
+}
+
+void VoiceRecorder::onSilenceCheck()
+{
+    if (!m_active)
+        return;
+    const qint64 elapsed = m_elapsed.elapsed();
+
+    // Nessuna voce dopo 12 s: la cattura viene scartata (failed "no-speech").
+    if (!m_hasVoice && elapsed >= 12000) {
+        qDebug() << "[voice] nessuna voce rilevata, stop";
+        finalize();
+        return;
+    }
+
+    // Stop automatico: almeno un frammento di voce udito e pausa oltre soglia.
+    if (m_hasVoice
+            && elapsed - m_lastVoiceMs >= qint64(m_settings->vadSilenceMs())
+            && elapsed >= 800) {
+        qDebug() << "[voice] stop automatico: pausa di" << (elapsed - m_lastVoiceMs)
+                 << "ms dopo voce a" << m_lastVoiceMs << "ms";
+        finalize();
+        return;
+    }
+
+    // Tetto di sicurezza anti-registrazione infinita.
+    if (elapsed >= 60000) {
+        qDebug() << "[voice] stop automatico: durata massima raggiunta";
+        finalize();
+    }
+}
+
+void VoiceRecorder::finalize()
+{
+    m_silenceTimer->stop();
+
+    if (m_device) {
+        m_pcm.append(m_device->readAll());
+        m_device = nullptr;
+    }
+    if (m_input) {
+        m_input->stop();
+        m_input->deleteLater();
+        m_input = nullptr;
+    }
+
+    m_active = false;
+    m_level = 0.0;
+    emit activeChanged();
+    emit levelChanged();
+
+    if (!m_hasVoice) {
+        m_pcm.clear();
+        emit failed(QStringLiteral("no-speech"));
+        return;
+    }
+
+    const QString path = makeWavPath();
+    const int bytes = m_pcm.size();
+    if (!writeWav(path)) {
+        m_pcm.clear();
+        emit failed(tr("Could not write the audio file"));
+        return;
+    }
+    m_pcm.clear();
+    qDebug() << "[voice] WAV pronto:" << path << bytes << "byte";
+    emit finished(path);
+}
+
+QString VoiceRecorder::makeWavPath() const
+{
+    const QString dir = QStandardPaths::writableLocation(QStandardPaths::AppDataLocation)
+            + QStringLiteral("/voice");
+    QDir().mkpath(dir);
+    return dir + QStringLiteral("/v_")
+            + QString::number(QDateTime::currentMSecsSinceEpoch())
+            + QStringLiteral(".wav");
+}
+
+bool VoiceRecorder::writeWav(const QString &path)
+{
+    QFile f(path);
+    if (!f.open(QIODevice::WriteOnly))
+        return false;
+
+    const quint16 channels = quint16(m_format.channelCount());
+    const quint32 sampleRate = quint32(m_format.sampleRate());
+    const quint16 bits = quint16(m_format.sampleSize());
+    const quint32 dataSize = quint32(m_pcm.size());
+    const quint32 byteRate = sampleRate * channels * bits / 8;
+    const quint16 blockAlign = quint16(channels * bits / 8);
+
+    QByteArray header;
+    header.reserve(44);
+    header.append("RIFF");
+    append32(header, 36 + dataSize);
+    header.append("WAVE");
+    header.append("fmt ");
+    append32(header, 16);
+    append16(header, 1); // PCM
+    append16(header, channels);
+    append32(header, sampleRate);
+    append32(header, byteRate);
+    append16(header, blockAlign);
+    append16(header, bits);
+    header.append("data");
+    append32(header, dataSize);
+
+    f.write(header);
+    f.write(m_pcm);
+    f.close();
+    return true;
+}
@@ -0,0 +1,73 @@
+#ifndef VOICERECORDER_H
+#define VOICERECORDER_H
+
+#include <QAudioFormat>
+#include <QByteArray>
+#include <QElapsedTimer>
+#include <QObject>
+#include <QString>
+
+class QAudioInput;
+class QIODevice;
+class QTimer;
+class Settings;
+
+// Registratore per la modalità dialogo (voce continua).
+//
+// Differenze rispetto a Recorder (dettatura manuale, QAudioRecorder/OGG):
+// - usa QAudioInput con PCM 16 bit mono per analizzare il livello del
+//   segnale in tempo reale;
+// - ferma la registrazione DA SOLO dopo un periodo di silenzio (pausa e
+//   soglia configurabili nelle impostazioni) oppure al tetto massimo;
+// - produce un WAV che il server trascrive come gli altri formati.
+//
+// finished(path) arriva per stop automatico o manuale; se non è stata
+// rilevata voce emette failed("no-speech") (il WAV viene scartato).
+// Il segnale listening() conferma l'avvio della cattura.
+class VoiceRecorder : public QObject
+{
+    Q_OBJECT
+    Q_PROPERTY(bool active READ active NOTIFY activeChanged)
+    Q_PROPERTY(qreal level READ level NOTIFY levelChanged)
+
+public:
+    explicit VoiceRecorder(Settings *settings, QObject *parent = nullptr);
+
+    bool active() const { return m_active; }
+    qreal level() const { return m_level; }
+
+    Q_INVOKABLE void start();
+    Q_INVOKABLE void stop();
+
+signals:
+    void activeChanged();
+    void levelChanged();
+    void listening();
+    void finished(const QString &wavPath);
+    void failed(const QString &error);
+
+private slots:
+    void onReadyRead();
+    void onSilenceCheck();
+
+private:
+    void finalize();
+    QString makeWavPath() const;
+    bool writeWav(const QString &path);
+
+    Settings *m_settings;
+    QAudioInput *m_input;
+    QIODevice *m_device;
+    QTimer *m_silenceTimer;
+    QAudioFormat m_format;
+
+    QByteArray m_pcm;
+    QElapsedTimer m_elapsed;
+    qint64 m_lastVoiceMs;
+    qint64 m_lastLevelEmitMs;
+    bool m_hasVoice;
+    bool m_active;
+    qreal m_level;
+};
+
+#endif // VOICERECORDER_H
@@ -21,6 +21,7 @@ SOURCES += \
    $$SRC/chatmodel.cpp \
    $$SRC/sessionsmodel.cpp \
    $$SRC/recorder.cpp \
+    $$SRC/voicerecorder.cpp \
    $$SRC/player.cpp \
    core_test.cpp

@@ -32,4 +33,5 @@ HEADERS += \
    $$SRC/chatmodel.h \
    $$SRC/sessionsmodel.h \
    $$SRC/recorder.h \
+    $$SRC/voicerecorder.h \
    $$SRC/player.h