hl-build-voice-app

---

name: hl-build-voice-app description: Build voice apps for all Hailo accelerators: Whisper STT on Hailo-8/8L/10H + optional Piper TTS on CPU.

Skill: Build Voice Application

Build voice apps for all Hailo accelerators: Whisper STT on Hailo-8/8L/10H + optional Piper TTS on CPU.

When This Skill Is Loaded

• User wants speech input or speech output in a Hailo app
• User mentions: voice, speech, Whisper, TTS, microphone, STT, speak, listen
• User wants to add voice to an existing LLM or VLM app
• User wants speech recognition on Hailo-8 or Hailo-8L

Hardware Compatibility

Feature	Hailo-8/8L	Hailo-10H
STT (Whisper)	✓ via InferModel (encoder+decoder HEFs)	✓ via Speech2Text (genai API)
LLM on device	✘	✓ via hailo_platform.genai.LLM
VLM on device	✘	✓ via Backend (VLM chat)
TTS (Piper)	✓ CPU	✓ CPU
Full voice assistant	STT + CPU LLM + TTS	STT + on-device LLM + TTS

Reference Implementations

Study these:

• hailo_apps/python/gen_ai_apps/voice_assistant/ — Full voice + LLM assistant (Hailo-10H)
• hailo_apps/python/gen_ai_apps/simple_whisper_chat/ — Simple STT example (Hailo-10H)
• hailo_apps/python/standalone_apps/speech_recognition/ — STT for all Hailo devices (8/8L/10H) using InferModel API:
  • speech_recognition.py — Main app: mic recording, audio preprocessing, transcription loop
  • whisper_pipeline.py — WhisperPipeline class: encoder+decoder inference via InferModel
  • audio_utils.py — Audio recording, mel spectrogram, file I/O
  • postprocessing.py — Repetition penalty and token decoding
• hailo_apps/python/gen_ai_apps/gen_ai_utils/voice_processing/ — Voice utilities (Hailo-10H):
  • speech_to_text.py — SpeechToTextProcessor (Whisper via genai API)
  • text_to_speech.py — TextToSpeechProcessor (Piper on CPU)
  • audio_recorder.py — AudioRecorder (microphone capture)
  • vad.py — Voice Activity Detection
  • interaction.py — VoiceInteractionManager (high-level orchestrator)

Build Process

Step 1: Create App Directory

Create the app directory:

hailo_apps/python/<type>/<app_name>/
├── app.yaml              # App manifest (type: gen_ai)
├── run.sh                # Launch wrapper
├── __init__.py
├── <app_name>.py         # Main app
└── README.md             # Usage documentation (REQUIRED — never skip)

Create app.yaml with type: gen_ai and run.sh wrapper. Do NOT register in defines.py or resources_config.yaml.

Step 2: Build Main App (Hailo-10H: Voice + LLM)

import signal
import threading
from contextlib import redirect_stderr
from io import StringIO

from hailo_platform import VDevice
from hailo_platform.genai import LLM

from hailo_apps.python.core.common.hailo_logger import get_logger
from hailo_apps.python.core.common.core import resolve_hef_path
from hailo_apps.python.core.common.parser import get_standalone_parser
from hailo_apps.python.core.common.defines import (
    SHARED_VDEVICE_GROUP_ID,
    HAILO10H_ARCH,
)

logger = get_logger(__name__)

from hailo_apps.python.gen_ai_apps.gen_ai_utils.voice_processing.speech_to_text import SpeechToTextProcessor
from hailo_apps.python.gen_ai_apps.gen_ai_utils.voice_processing.text_to_speech import TextToSpeechProcessor
from hailo_apps.python.gen_ai_apps.gen_ai_utils.voice_processing.interaction import VoiceInteractionManager
from hailo_apps.python.gen_ai_apps.gen_ai_utils.voice_processing.vad import add_vad_args

APP_NAME = "my_voice_app"

logger = get_logger(__name__)

APP_NAME = MY_VOICE_APP
SYSTEM_PROMPT = "You are a helpful voice assistant. Keep responses concise and natural."


def main():
    parser = get_standalone_parser()
    parser.add_argument("--no-tts", action="store_true", help="Disable TTS (text only)")
    parser.add_argument("--system-prompt", type=str, default=SYSTEM_PROMPT)
    add_vad_args(parser)
    args = parser.parse_args()

    abort_event = threading.Event()
    signal.signal(signal.SIGINT, lambda s, f: abort_event.set())

    # VDevice
    params = VDevice.create_params()
    params.group_id = SHARED_VDEVICE_GROUP_ID
    vdevice = VDevice(params)

    # STT (Whisper on Hailo)
    whisper_hef = resolve_hef_path(None, "whisper", arch=HAILO10H_ARCH)
    with redirect_stderr(StringIO()):  # Suppress ALSA noise
        stt = SpeechToTextProcessor(vdevice, str(whisper_hef))

    # LLM (on Hailo)
    llm_hef = resolve_hef_path(args.hef_path, APP_NAME, arch=HAILO10H_ARCH)
    llm = LLM(vdevice, str(llm_hef))

    # TTS (Piper on CPU)
    tts = None if args.no_tts else TextToSpeechProcessor()

    # Voice interaction manager
    vim = VoiceInteractionManager(stt, tts, abort_event)

    logger.info("Voice assistant ready. Speak into your microphone.")
    print("Voice assistant ready. Press Ctrl+C to quit.\n")

    try:
        while not abort_event.is_set():
            # Listen for speech
            user_text = vim.listen()
            if not user_text or abort_event.is_set():
                continue

            logger.info("User said: %s", user_text)
            print(f"You: {user_text}")

            # Generate response
            prompt = [
                {"role": "system", "content": [{"type": "text", "text": args.system_prompt}]},
                {"role": "user", "content": [{"type": "text", "text": user_text}]},
            ]
            response = llm.generate_all(
                prompt=prompt, temperature=0.1, max_generated_tokens=150
            )
            llm.clear_context()

            print(f"Assistant: {response}\n")

            # Speak response
            if tts and not abort_event.is_set():
                vim.speak(response)
    finally:
        llm.release()
        vdevice.release()
        logger.info("Cleanup complete")


if __name__ == "__main__":
    main()

Step 3: Validate (Hailo-10H)

python3 .github/scripts/validate_app.py hailo_apps/python/gen_ai_apps/my_voice_app --smoke-test

Step 2b: Build Main App (Hailo-8/8L: Speech Recognition)

For Hailo-8/8L, use the InferModel API with separate encoder + decoder HEFs:

import signal
import sys
import time

import numpy as np

from hailo_apps.python.core.common.toolbox import resolve_arch
from hailo_apps.python.core.common.hailo_logger import get_logger
from hailo_apps.python.core.common.core import resolve_hef_paths
from hailo_apps.python.core.common.parser import get_standalone_parser
from hailo_apps.python.core.common.defines import (
    WHISPER_H8_APP, RESOURCES_ROOT_PATH_DEFAULT, RESOURCES_NPY_DIR_NAME,
)

logger = get_logger(__name__)

APP_NAME = "my_speech_app"


def main():
    parser = get_standalone_parser()
    parser.add_argument("--audio", type=str, help="Audio file to transcribe (skip mic)")
    parser.add_argument("--variant", default="base", choices=["base", "tiny", "tiny.en"],
                        help="Whisper model variant")
    parser.add_argument("--duration", type=int, default=10, help="Max recording seconds")
    args = parser.parse_args()

    arch = resolve_arch(args.arch)  # Auto-detect or --arch flag

    # Resolve encoder + decoder HEFs (plural — Whisper uses two separate models)
    hef_paths = resolve_hef_paths(None, WHISPER_H8_APP, arch=arch)
    encoder_path = str(hef_paths["encoder"])
    decoder_path = str(hef_paths["decoder"])

    # Decoder assets directory
    npy_dir = str(Path(RESOURCES_ROOT_PATH_DEFAULT) / RESOURCES_NPY_DIR_NAME)

    # WhisperPipeline — uses InferModel API, works on ALL Hailo devices
    from .whisper_pipeline import WhisperPipeline
    pipeline = WhisperPipeline(
        encoder_path=encoder_path,
        decoder_path=decoder_path,
        variant=args.variant,
        npy_dir=npy_dir,
        add_embed=(arch != "hailo10h"),  # H8/8L need host-side Add operator
    )

    if args.audio:
        # Single file transcription
        mel = preprocess_audio_file(args.audio, pipeline.get_chunk_length())
        pipeline.send_data(mel)
        text = pipeline.get_transcription()
        print(f"Transcription: {text}")
    else:
        # Interactive mic loop
        logger.info("Ready. Press Enter to start recording, Enter to stop. 'q' to quit.")
        while True:
            cmd = input("Press Enter to record (q to quit): ").strip()
            if cmd.lower() == "q":
                break
            audio = record_from_mic(duration=args.duration)
            mel = audio_to_mel(audio, pipeline.get_chunk_length())
            pipeline.send_data(mel)
            text = pipeline.get_transcription()
            print(f"Transcription: {text}")

    pipeline.stop()
    logger.info("Done")


if __name__ == "__main__":
    main()

Key differences from H10 pattern:

• Uses resolve_hef_paths() (plural) — returns dict with encoder and decoder keys
• Uses WhisperPipeline with InferModel API — NOT SpeechToTextProcessor
• add_embed flag: True for H8/8L (host-side Add operator), False for H10
• No LLM/VLM on device — inference is CPU or external service
• Dependencies: pip install -e ".[speech-rec]" (torch, transformers, sounddevice, scipy)

Step 3b: Validate (Hailo-8/8L)

python3 .github/scripts/validate_app.py hailo_apps/python/standalone_apps/my_speech_app --smoke-test

Critical Conventions

1. STT on Hailo, TTS on CPU: Never reverse this — Whisper needs the accelerator, Piper is CPU-only
2. Two STT APIs: SpeechToTextProcessor (genai) for H10, WhisperPipeline (InferModel) for H8/8L/10H
3. Architecture detection: Use detect_hailo_arch() or --arch flag — never hardcode
4. ALSA noise: Wrap audio init with redirect_stderr(StringIO())
5. Abort event: threading.Event() for interrupting generation and speech
6. Init order: VDevice → STT → LLM (H10 only) → TTS → VoiceInteractionManager
7. Cleanup order: Reverse of init, in finally block
8. VAD args: Always use add_vad_args(parser) for --vad, --vad-aggressiveness, --vad-energy-threshold
9. --no-tts: Always support text-only mode as fallback
10. H8/8L HEF resolution: Use resolve_hef_paths() (plural) — Whisper needs encoder + decoder HEFs
11. H10 HEF resolution: Use resolve_hef_path() (singular) — single HEF per model
12. Dependencies: H10 voice requires [gen-ai] extras; H8/8L STT requires [speech-rec] extras

Adding Voice to an Existing App

To add voice to any existing LLM/VLM app:

1. Import voice utilities
2. Add --voice and --no-tts CLI flags + add_vad_args(parser)
3. Init STT + TTS alongside existing model
4. Replace input() with vim.listen() when --voice is set
5. Add vim.speak(response) after generating output
6. Add abort_event for interruption support