Files
hermes-agent/tools/voice_mode_transcript.py

111 lines
5.4 KiB
Python

"""Pure-text helpers for voice mode: Whisper hallucination filter, voice-chat
stop phrases, and the TTS self-echo guard. No audio dependencies."""
import difflib
import re
from contextlib import suppress
from typing import Optional
def _voice_config() -> dict:
"""``voice`` section of config.yaml, or ``{}`` when missing, malformed, or the
config system can't be imported (broken config mid-install)."""
with suppress(Exception):
from hermes_cli.config import load_config
voice_cfg = load_config().get("voice", {})
return voice_cfg if isinstance(voice_cfg, dict) else {}
return {}
# Whisper commonly hallucinates these phrases on silent/near-silent audio
# (matched with trailing '.'/'!' stripped, so the bare form suffices).
WHISPER_HALLUCINATIONS = {
"thank you", "thanks for watching", "subscribe to my channel", "like and subscribe",
"please subscribe", "thank you for watching", "bye", "you", "the end",
# Non-English hallucinations (common on silence)
"продолжение следует", "sous-titres", "sous-titres réalisés par la communauté d'amara.org",
"sottotitoli creati dalla comunità amara.org", "untertitel von stephanie geiges",
"amara.org", "www.mooji.org", "ご視聴ありがとうございました",
}
# Repetitive hallucinations (e.g. "Thank you. Thank you. Thank you.")
_HALLUCINATION_REPEAT_RE = re.compile(r'^(?:thank you|thanks|bye|you|ok|okay|the end|\.|\s|,|!)+$',
flags=re.IGNORECASE)
def is_whisper_hallucination(transcript: str) -> bool:
"""Check if a transcript is a known Whisper hallucination on silence."""
cleaned = transcript.strip().lower()
return (not cleaned or cleaned.rstrip('.!') in WHISPER_HALLUCINATIONS
or bool(_HALLUCINATION_REPEAT_RE.match(cleaned)))
DEFAULT_VOICE_STOP_PHRASES = ("stop",)
def _load_voice_stop_phrases() -> tuple:
"""Configured ``voice.stop_phrases`` (default ``("stop",)``); an empty tuple disables
the feature. Malformed config (dict, list of non-strings) falls back to the default
rather than crashing the voice loop."""
with suppress(Exception):
raw = _voice_config().get("stop_phrases", DEFAULT_VOICE_STOP_PHRASES)
if isinstance(raw, str):
raw = [raw]
if isinstance(raw, (list, tuple)):
return tuple(str(p).strip().lower() for p in raw if isinstance(p, (str, int, float)) and str(p).strip())
return DEFAULT_VOICE_STOP_PHRASES
def is_voice_stop_phrase(transcript: str, stop_phrases: Optional[tuple] = None) -> bool:
"""True when *transcript* is EXACTLY a configured stop phrase. Deliberately strict: the whole
utterance — lowercased, surrounding punctuation stripped — must equal a phrase, so "stop doing
that and try again" still reaches the agent. ``voice.stop_phrases: []`` disables."""
cleaned = transcript.strip().lower().strip(".,!?;: \t\n\"'") if transcript else ""
return bool(cleaned) and cleaned in (_load_voice_stop_phrases() if stop_phrases is None else stop_phrases)
# Similarity ratio (difflib.SequenceMatcher) above which a playback-phase barge transcript
# is treated as a self-capture of Hermes' own TTS: the full-duplex listener has no echo
# cancellation, so speaker bleed can be transcribed near-verbatim (TTS -> STT -> TTS loop).
DEFAULT_TTS_ECHO_SIMILARITY_THRESHOLD = 0.6
# Minimum normalized-transcript length before the sliding-window fallback runs. Below
# this a genuine one-word barge-in ("yes") landing verbatim inside a longer reply would
# score a trivial 1.0; a real self-capture spans pre-roll plus time-to-silence, so it is longer.
# See #75792.
MIN_FRAGMENT_LENGTH_FOR_ECHO = 10
def _normalize_for_echo_compare(text: str) -> str:
return re.sub(r"\s+", " ", text).strip().lower()
def is_tts_echo(transcript: str, spoken_text: str,
threshold: float = DEFAULT_TTS_ECHO_SIMILARITY_THRESHOLD) -> bool:
"""True when *transcript* looks like a self-capture of *spoken_text*. Character-level similarity
(language-agnostic): a genuine interjection rarely matches Hermes' own words, so a high ratio signals
speaker-bleed (fail-closed guard for the playback-phase listener). Playback capture spans only pre-roll
plus time-to-silence, so for long replies the transcript is a FRAGMENT and the whole-string ratio dilutes
toward 0; when it misses, a transcript-sized window slides across `spoken_text`. Transcripts shorter than
`MIN_FRAGMENT_LENGTH_FOR_ECHO` skip the fallback (a short interjection trivially matches a short window)."""
a, b = _normalize_for_echo_compare(transcript or ""), _normalize_for_echo_compare(spoken_text or "")
if not a or not b:
return False
def _similar(x: str, y: str) -> bool:
return difflib.SequenceMatcher(None, x, y).ratio() >= threshold
if _similar(a, b):
return True
if len(a) < MIN_FRAGMENT_LENGTH_FOR_ECHO or len(a) >= len(b):
return False
return any(_similar(a, b[start : start + len(a)]) for start in range(0, len(b) - len(a) + 1))
def voice_stop_hint() -> str:
"""One-line 'Say "stop" to end the voice chat.' hint for voice-mode start, using the first
``voice.stop_phrases`` entry ("" when disabled). Every surface announcing voice-mode start
(CLI, TUI, desktop) uses this one owner instead of hardcoding the wording."""
phrases = _load_voice_stop_phrases()
return f'Say "{phrases[0]}" to end the voice chat.' if phrases else ""