feat: add --language flag for non-English TTS voices (Dutch support)

- Add LANGUAGE default (en) to shell.functions - setup_python_venv downloads Dutch ONNX voices (pim, ronnie, nathalie) - wake_word_sample_generator uses multiple --model flags for single-speaker voices, cycling between them for variety - train_wake_word accepts and passes --language through the pipeline - recorder_server.py accepts language in session API - Web UI adds language dropdown (English/Dutch)
2026-06-12 20:10:19 -06:00 · 2026-02-25 09:57:46 +01:00
parent 747822e856
commit d5dcfbf5f1
6 changed files with 113 additions and 18 deletions
--- a/cli/setup_python_venv
+++ b/cli/setup_python_venv
@@ -157,6 +157,10 @@ pip_install -e "${PSG}"
 git -C tools/piper-sample-generator clean -fd &>/dev/null

 MODELS_DIR="${PSG}/models"
+VOICES_DIR="${PSG}/voices"
+mkdir -p "${VOICES_DIR}"
+
+# --- English generator model (multi-speaker, used with --language=en) ---
 MODEL_NAME="en_US-libritts_r-medium.pt"
 MODEL_FILE="${MODELS_DIR}/${MODEL_NAME}"
 MODEL_URL="https://github.com/rhasspy/piper-sample-generator/releases/download/v2.0.0/${MODEL_NAME}"
@@ -170,6 +174,30 @@ if [ ! -f "${MODEL_FILE}.json" ] ; then
    curl -sfL "${MODEL_URL}.json" -o "${MODEL_FILE}.json"
 fi

+# --- Dutch ONNX voices (single-speaker, used with --language=nl) ---
+# Working Dutch voices: pim, ronnie (nl_NL) and nathalie (nl_BE).
+# nl_NL-mls-medium is intentionally excluded (known Piper issue: outputs gibberish).
+HF_VOICES="https://huggingface.co/rhasspy/piper-voices/resolve/main"
+declare -a NL_VOICES=(
+    "nl/nl_NL/pim/medium/nl_NL-pim-medium"
+    "nl/nl_NL/ronnie/medium/nl_NL-ronnie-medium"
+    "nl/nl_BE/nathalie/medium/nl_BE-nathalie-medium"
+)
+echo "   ===== Checking Dutch Piper voices ====="
+for voice_path in "${NL_VOICES[@]}" ; do
+    voice_name="$(basename "${voice_path}")"
+    onnx_file="${VOICES_DIR}/${voice_name}.onnx"
+    json_file="${VOICES_DIR}/${voice_name}.onnx.json"
+    if [ ! -f "${onnx_file}" ] ; then
+        echo "      Downloading ${voice_name}.onnx"
+        curl -sfL "${HF_VOICES}/${voice_path}.onnx?download=true" -o "${onnx_file}"
+    fi
+    if [ ! -f "${json_file}" ] ; then
+        echo "      Downloading ${voice_name}.onnx.json"
+        curl -sfL "${HF_VOICES}/${voice_path}.onnx.json?download=true" -o "${json_file}"
+    fi
+done
+
 ${GPU} && onnxgpu='-gpu[cuda]' || onnxgpu=""
 echo "   ===== Installing onnxruntime${onnxgpu} ====="
 pip_install "onnxruntime${onnxgpu}>=1.16.0"
--- a/cli/shell.functions
+++ b/cli/shell.functions
@@ -11,12 +11,14 @@ fi
 DEFAULT_SAMPLES=50000
 DEFAULT_BATCH_SIZE=100
 DEFAULT_TRAINING_STEPS=40000
+DEFAULT_LANGUAGE=en

 [ -f "${DATA_DIR}/.defaults.env" ] && source "${DATA_DIR}/.defaults.env" || :

 : "${SAMPLES:=${DEFAULT_SAMPLES}}"
 : "${BATCH_SIZE:=${DEFAULT_BATCH_SIZE}}"
 : "${TRAINING_STEPS:=${DEFAULT_TRAINING_STEPS}}"
+: "${LANGUAGE:=${DEFAULT_LANGUAGE}}"
 : "${CLEANUP_WORK_DIR:=false}"
 : "${CLEANUP_ARCHIVES:=false}"
 : "${CLEANUP_INTERMEDIATE_FILES:=false}"
--- a/cli/wake_word_sample_generator
+++ b/cli/wake_word_sample_generator
@@ -4,7 +4,7 @@ set -e
 PROGPATH=$(realpath "$0")
 PROGDIR=$(dirname "${PROGPATH}")

-KNOWN_ARGS=( samples batch-size data-dir )
+KNOWN_ARGS=( samples batch-size data-dir language )
 source "${PROGDIR}/shell.functions"
 WAKE_WORD="${POSITIONAL_ARGS[0]}"

@@ -15,7 +15,8 @@ fi

 if [ "${HELP}" == "true" ] || [ -z "${WAKE_WORD}" ] ; then
    cat <<EOF >&2
-Usage: $0 [ --samples=<samples> ] [ --batch-size=<batch_size> ] <wake_word>
+Usage: $0 [ --samples=<samples> ] [ --batch-size=<batch_size> ]
+          [ --language=<lang> ] <wake_word>

 --samples:            The number of samples to generate for the wake word.
                      Default: ${DEFAULT_SAMPLES}
@@ -24,6 +25,12 @@ Usage: $0 [ --samples=<samples> ] [ --batch-size=<batch_size> ] <wake_word>
                      samples, the more memory is needed.
                      Default: ${DEFAULT_BATCH_SIZE}

+--language:           Language for TTS voice selection.
+                      "en" uses the multi-speaker LibriTTS-R generator.
+                      Other languages (e.g. "nl") use single-speaker ONNX
+                      voices and cycle between them for variety.
+                      Default: ${DEFAULT_LANGUAGE}
+
 <wake_word>           The word to generate samples for.
                      Required.

@@ -40,21 +47,60 @@ cd "${WORK_DIR}"

 PSG="${DATA_DIR}/tools/piper-sample-generator"
 MODELS_DIR="${PSG}/models"
-MODEL_NAME=en_US-libritts_r-medium.pt
-MODEL_FILE="${MODELS_DIR}/${MODEL_NAME}"
+VOICES_DIR="${PSG}/voices"
 SAMPLES_DIR="${WORK_DIR}/wake_word_samples"

 mkdir -p "${SAMPLES_DIR}" || :

+# ---------------------------------------------------------------------------
+# Build the --model argument(s) based on language
+# ---------------------------------------------------------------------------
+declare -a MODEL_ARGS=()
+MODEL_TAG=""
+
+if [ "${LANGUAGE}" == "en" ] ; then
+    # English: use the multi-speaker LibriTTS-R generator (.pt)
+    MODEL_NAME="en_US-libritts_r-medium.pt"
+    MODEL_FILE="${MODELS_DIR}/${MODEL_NAME}"
+    if [ ! -f "${MODEL_FILE}" ] ; then
+        echo "ERROR: English model ${MODEL_FILE} not found. Run setup_python_venv first." >&2
+        exit 1
+    fi
+    MODEL_ARGS=( --model "${MODEL_FILE}" )
+    MODEL_TAG="${MODEL_NAME}"
+else
+    # Non-English: find all ONNX voices matching the language prefix
+    # e.g. LANGUAGE=nl matches nl_NL-pim-medium.onnx, nl_BE-nathalie-medium.onnx, etc.
+    shopt -s nullglob
+    voice_files=( "${VOICES_DIR}/${LANGUAGE}"_*.onnx )
+    shopt -u nullglob
+
+    if [ ${#voice_files[@]} -eq 0 ] ; then
+        echo "ERROR: No ONNX voice files found for language '${LANGUAGE}' in ${VOICES_DIR}/" >&2
+        echo "       Expected files matching: ${LANGUAGE}_*.onnx" >&2
+        echo "       Run setup_python_venv to download voice models." >&2
+        exit 1
+    fi
+
+    echo "   Using ${#voice_files[@]} voice(s) for language '${LANGUAGE}':"
+    MODEL_TAG="${LANGUAGE}"
+    for vf in "${voice_files[@]}" ; do
+        vname="$(basename "${vf}")"
+        echo "      - ${vname}"
+        MODEL_ARGS+=( --model "${vf}" )
+        MODEL_TAG="${MODEL_TAG}+${vname}"
+    done
+fi
+
 REGENERATE=false

 if [ "${SAMPLES}" -eq 1 ] ; then
-    echo "===== Generating ${SAMPLES} sample of '${WAKE_WORD}' ====="
-    wake_word_filename="${WAKE_WORD//[ \`~\!\$&*\(\)\{\}\[\]\|\;\'\"<>.?\/]/_}"
+    echo "===== Generating ${SAMPLES} sample of '${WAKE_WORD}' (language=${LANGUAGE}) ====="
+    wake_word_filename="${WAKE_WORD//[ \`~\!@#\$%^&*\(\)\{\}\[\]\|\;\'\"<>.?\/]/_}"

    mkdir -p "${WORK_DIR}/test_sample" || :
    "${PSG}/generate_samples.py" "${WAKE_WORD}" \
-        --model "${MODEL_FILE}" \
+        "${MODEL_ARGS[@]}" \
        --max-samples ${SAMPLES} \
        --batch-size ${BATCH_SIZE} \
        --output-dir "${WORK_DIR}/test_sample" \
@@ -65,9 +111,9 @@ if [ "${SAMPLES}" -eq 1 ] ; then
    exit 0
 fi

-grep -q "${WAKE_WORD}:${SAMPLES}:${MODEL_NAME}" "${WORK_DIR}/last_wake_word" &>/dev/null || REGENERATE=true
+grep -q "${WAKE_WORD}:${SAMPLES}:${MODEL_TAG}" "${WORK_DIR}/last_wake_word" &>/dev/null || REGENERATE=true

-# Double check that the number of existing samples matches SAMPLES"
+# Double check that the number of existing samples matches SAMPLES
 existing_samples=$(find "${SAMPLES_DIR}" -name '*.wav' | wc -l)
 [ "${existing_samples}" -eq "${SAMPLES}" ] || REGENERATE=true

@@ -79,7 +125,7 @@ if ! ${REGENERATE} ; then
    exit 0
 fi

-echo -e "\n===== Generating ${SAMPLES} wake word samples in batches of ${BATCH_SIZE}  ====="
+echo -e "\n===== Generating ${SAMPLES} wake word samples in batches of ${BATCH_SIZE} (language=${LANGUAGE}) ====="
 export TF_CPP_MIN_LOG_LEVEL=9
 export TF_FORCE_GPU_ALLOW_GROWTH=true
 export TF_GPU_ALLOCATOR=cuda_malloc_async
@@ -93,7 +139,7 @@ echo "   Generating samples"
 rm -rf "${SAMPLES_DIR}" || :
 mkdir -p "${SAMPLES_DIR}" || :
 "${PSG}/generate_samples.py" "${WAKE_WORD}" \
-    --model "${MODEL_FILE}" \
+    "${MODEL_ARGS[@]}" \
    --max-samples ${SAMPLES} \
    --batch-size ${BATCH_SIZE} \
    --output-dir "${SAMPLES_DIR}" 2>&1 | sed -r -e "s/(DEBUG|INFO):__main__:/      /g"
@@ -103,8 +149,7 @@ if [ "${generated_files}" -ne "${SAMPLES}" ] ; then
    echo "ERROR:  only generated ${generated_files} files" >&2
    exit 1
 fi
-END_TS=$(date +%s.%N)
-echo "${WAKE_WORD}:${SAMPLES}:${MODEL_NAME}" > "${WORK_DIR}/last_wake_word"
+echo "${WAKE_WORD}:${SAMPLES}:${MODEL_TAG}" > "${WORK_DIR}/last_wake_word"
 echo
 END_TS=$EPOCHSECONDS
 print_elapsed_time "${START_TS}" "${END_TS}" "Generated ${SAMPLES} wake word samples."
--- a/recorder_server.py
+++ b/recorder_server.py
@@ -35,6 +35,8 @@ TRAIN_CMD = os.environ.get(
    f"source '{DATA_DIR}/.venv/bin/activate' && train_wake_word --data-dir '{DATA_DIR}'"
 )

+DEFAULT_LANGUAGE = os.environ.get("MWW_LANGUAGE", "en")
+
 TAKES_PER_SPEAKER_DEFAULT = int(os.environ.get("REC_TAKES_PER_SPEAKER", "10"))
 SPEAKERS_TOTAL_DEFAULT = int(os.environ.get("REC_SPEAKERS_TOTAL", "1"))

@@ -60,6 +62,7 @@ def safe_name(raw: str) -> str:
 STATE: Dict[str, Any] = {
    "raw_phrase": None,
    "safe_word": None,
+    "language": DEFAULT_LANGUAGE,

    "speakers_total": SPEAKERS_TOTAL_DEFAULT,
    "takes_per_speaker": TAKES_PER_SPEAKER_DEFAULT,
@@ -372,6 +375,7 @@ def _normalize_output_artifacts(safe_word: str, log_path: Path) -> None:
 def _run_training_background(safe_word: str, allow_no_personal: bool):
    with STATE_LOCK:
        raw_phrase = STATE.get("raw_phrase") or ""
+        language = STATE.get("language") or DEFAULT_LANGUAGE

    wake_word_title = _title_from_phrase(raw_phrase)

@@ -403,9 +407,9 @@ def _run_training_background(safe_word: str, allow_no_personal: bool):
        _ensure_training_datasets(log_path)

        if wake_word_title:
-            cmd_str = f"{TRAIN_CMD} '{safe_word}' '{wake_word_title}'"
+            cmd_str = f"{TRAIN_CMD} --language='{language}' '{safe_word}' '{wake_word_title}'"
        else:
-            cmd_str = f"{TRAIN_CMD} '{safe_word}'"
+            cmd_str = f"{TRAIN_CMD} --language='{language}' '{safe_word}'"

        env = os.environ.copy()
        env["MWW_ALLOW_NO_PERSONAL"] = "true" if allow_no_personal else "false"
@@ -470,6 +474,7 @@ def start_session(payload: Dict[str, Any]):

    speakers_total = int(payload.get("speakers_total") or SPEAKERS_TOTAL_DEFAULT)
    takes_per_speaker = int(payload.get("takes_per_speaker") or TAKES_PER_SPEAKER_DEFAULT)
+    language = (payload.get("language") or DEFAULT_LANGUAGE).strip().lower()

    speakers_total = max(1, min(10, speakers_total))
    takes_per_speaker = max(1, min(50, takes_per_speaker))
@@ -477,6 +482,7 @@ def start_session(payload: Dict[str, Any]):
    with STATE_LOCK:
        STATE["raw_phrase"] = raw
        STATE["safe_word"] = safe
+        STATE["language"] = language
        STATE["speakers_total"] = speakers_total
        STATE["takes_per_speaker"] = takes_per_speaker
        STATE["takes_received"] = 0
@@ -491,6 +497,7 @@ def start_session(payload: Dict[str, Any]):
        "ok": True,
        "raw_phrase": raw,
        "safe_word": safe,
+        "language": language,
        "speakers_total": speakers_total,
        "takes_per_speaker": takes_per_speaker,
        "takes_total": speakers_total * takes_per_speaker,
@@ -506,6 +513,7 @@ def get_session():
            "ok": True,
            "raw_phrase": STATE["raw_phrase"],
            "safe_word": STATE["safe_word"],
+            "language": STATE["language"],
            "speakers_total": STATE["speakers_total"],
            "takes_per_speaker": STATE["takes_per_speaker"],
            "takes_received": STATE["takes_received"],
--- a/static/index.html
+++ b/static/index.html
@@ -178,6 +178,12 @@
      </div>

      <div class="row" style="margin-top:10px;">
+        <label class="muted">Language
+          <select id="language" style="padding:10px 12px;font-size:15px;border-radius:12px;border:1px solid rgba(255,255,255,0.12);background:rgba(0,0,0,0.35);color:var(--text);outline:none;">
+            <option value="en" selected>English (en)</option>
+            <option value="nl">Dutch (nl)</option>
+          </select>
+        </label>
        <label class="muted">Speakers
          <input id="speakersTotal" type="number" min="1" max="10" value="1" />
        </label>
@@ -727,13 +733,14 @@

    speakersTotal = parseInt($("speakersTotal").value || "1", 10);
    takesPerSpeaker = parseInt($("takesPerSpeaker").value || "10", 10);
+    const language = $("language").value || "en";

    try {
      setPill($("sessionPill"), "Starting…", "warn");
      const data = await api("/api/start_session", {
        method: "POST",
        headers: {"Content-Type":"application/json"},
-        body: JSON.stringify({ phrase, speakers_total: speakersTotal, takes_per_speaker: takesPerSpeaker })
+        body: JSON.stringify({ phrase, speakers_total: speakersTotal, takes_per_speaker: takesPerSpeaker, language })
      });

      session = data;
@@ -755,7 +762,7 @@

      await stopMicNow();

-      setPill($("sessionPill"), `Session: ${data.safe_word}`, "ok");
+      setPill($("sessionPill"), `Session: ${data.safe_word} (${data.language || "en"})`, "ok");
      $("beginBtn").disabled = false;
      $("resetBtn").disabled = false;
      $("trainBtn").disabled = false;
--- a/7
+++ b/7
@@ -5,7 +5,7 @@ PROGPATH=$(realpath "$0")
 PROGDIR=$(dirname "${PROGPATH}")
 CLIDIR="${PROGDIR}/cli"

-KNOWN_ARGS=( samples batch-size training-steps data-dir cleanup-work-dir )
+KNOWN_ARGS=( samples batch-size training-steps data-dir cleanup-work-dir language )
 source "${CLIDIR}/shell.functions"
 WAKE_WORD=${POSITIONAL_ARGS[0]}

@@ -18,6 +18,7 @@ if [ "${HELP}" == "true" ] || [ -z "${WAKE_WORD}" ] ; then
    cat <<EOF >&2
 Usage: train_wake_word [ --samples=<samples> ] [ --batch-size=<batch_size> ]
                       [ --training-steps=<steps> ] [ --cleanup-work-dir ]
+                       [ --language=<lang> ]
                       <wake_word> [ <wake_word_title> ]

 Options:
@@ -35,6 +36,9 @@ Options:
 --cleanup-work-dir:   Delete the /data/work directory after successful training.
                      Default: false

+--language:           Language for TTS voice selection (e.g. "en", "nl").
+                      Default: ${DEFAULT_LANGUAGE}
+
 <wake_word>           The word to train spelled phonetically.
                      Required.

@@ -88,6 +92,7 @@ export GRPC_VERBOSITY=ERROR
 "${CLIDIR}/wake_word_sample_generator" \
    --samples=${SAMPLES} \
    --batch-size=${BATCH_SIZE} \
+    --language="${LANGUAGE}" \
    --data-dir="${DATA_DIR}" "${WAKE_WORD}"

 POST_GEN_TS=$EPOCHSECONDS