thai-voice-command

Thai voice commands made simple. Give it text from any speech recognizer (sherpa-onnx, Whisper, Google, the browser's Web Speech API) and it will:

• detect a wake phrase even when the ASR misspells it (จาร์วิส, จาวิส, จาวิทย์ …)
• route the command to your function, ignoring spaces and polite particles (ครับ, ค่ะ)
• pull out arguments (ค้นหา แมว → "แมว")
• use fuzzy matching safely: unclear speech is ignored, never guessed
• cut microphone audio into utterances with a tiny pure-Python VAD

No required dependencies. Python 3.9+.

Extracted from JARVIS, a Thai desktop voice assistant, where real ASR mis-hearings were collected and tuned.

Install

pip install thai-voice-command

Quick start

from thai_voice_command import Router, WakeWord, JARVIS_ALIASES

router = Router(wake=WakeWord(JARVIS_ALIASES))

@router.command(phrases=["เปิดยูทูบ", "เปิด youtube"])
def youtube(match):
    return "opening YouTube"

@router.command(prefixes=["ค้นหา"])
def search(match):
    return f"searching for {match.arg}"

router.handle("จาร์วิส เปิด ยูทูบ ครับ")   # 'opening YouTube'
router.handle("จาวิทย์ ค้นหา Roblox")      # 'searching for Roblox'
router.handle("เปิดยูทูบ")                 # None (no wake phrase)
router.handle("จาร์วิส ทำกับข้าว")          # None (unknown, ignored)

How matching works

Checked in this order; the first hit wins.

Fuzzy matching accepts a command only when its score is at least min_score (default 0.78) and beats every other command by min_margin (default 0.065). So เปิดโน๊ตแพด is rejected when both "open Notepad" and "close Notepad" exist: the two scores (0.909 vs 0.857) are too close to be sure. For an assistant that launches programs, doing nothing is safer than doing the wrong thing.

Microphone

Segmenter turns 16-bit mono PCM blocks into whole utterances:

from thai_voice_command import Segmenter

seg = Segmenter(sample_rate=16000, block_samples=1280)
for block in mic_blocks:
    utterance = seg.feed(block)
    if utterance:
        text = my_asr(utterance)
        router.handle(text)

See examples/microphone_sherpa.py for a full offline Thai example.

Contributing

New wake-word aliases, Thai command phrases, and bug reports are very welcome. See CONTRIBUTING.md. Thai or English is fine.

License

MIT