meta: add Muse Voice Transcribe STT plugin

livekit-plugins/livekit-plugins-meta/README.md

@@ -0,0 +1,15 @@
+# Meta plugin for LiveKit Agents
+
+Support for speech-to-text with [Meta](https://dev.meta.ai/)'s Muse Voice Transcribe.
+
+See the [Meta integration docs](https://docs.livekit.io/agents/integrations/stt/meta/) for more information.
+
+## Installation
+
+```bash
+pip install livekit-plugins-meta
+```
+
+## Pre-requisites
+
+You'll need an API key from the Meta Model API. It can be set as an environment variable: `META_API_KEY`

livekit-plugins/livekit-plugins-meta/livekit/plugins/meta/__init__.py

@@ -0,0 +1,54 @@
+# Copyright 2026 LiveKit, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Meta plugin for LiveKit Agents
+
+Speech-to-text with Muse Voice Transcribe.
+See https://dev.meta.ai/docs/speech-to-text for more information.
+"""
+
+from .models import MuseEncoding, MuseMode, MuseModels, MusePartialMode
+from .stt import STT, SpeechStream
+from .version import __version__
+
+__all__ = [
+    "STT",
+    "MuseEncoding",
+    "MuseMode",
+    "MuseModels",
+    "MusePartialMode",
+    "SpeechStream",
+    "__version__",
+]
+
+from livekit.agents import Plugin
+
+from .log import logger
+
+
+class MetaPlugin(Plugin):
+    def __init__(self) -> None:
+        super().__init__(__name__, __version__, __package__, logger)
+
+
+Plugin.register_plugin(MetaPlugin())
+
+# Cleanup docs of unexported modules
+_module = dir()
+NOT_IN_ALL = [m for m in _module if m not in __all__]
+
+__pdoc__ = {}
+
+for n in NOT_IN_ALL:
+    __pdoc__[n] = False

livekit-plugins/livekit-plugins-meta/livekit/plugins/meta/_utils.py

@@ -0,0 +1,39 @@
+from __future__ import annotations
+
+import time
+from typing import TYPE_CHECKING, Generic, TypeVar
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+
+T = TypeVar("T")
+
+
+class PeriodicCollector(Generic[T]):
+    def __init__(self, callback: Callable[[T], None], *, duration: float) -> None:
+        """Accumulate values and hand the total to `callback` every `duration` seconds.
+
+        Args:
+            callback: Called with the accumulated total when `duration` elapses.
+            duration: Seconds between callback invocations.
+        """
+        self._duration = duration
+        self._callback = callback
+        self._last_flush_time = time.monotonic()
+        self._total: T | None = None
+
+    def push(self, value: T) -> None:
+        """Add a value to the accumulator."""
+        if self._total is None:
+            self._total = value
+        else:
+            self._total += value  # type: ignore[operator]
+        if time.monotonic() - self._last_flush_time >= self._duration:
+            self.flush()
+
+    def flush(self) -> None:
+        """Report the current total, if any, and start a new period."""
+        if self._total is not None:
+            self._callback(self._total)
+            self._total = None
+        self._last_flush_time = time.monotonic()

livekit-plugins/livekit-plugins-meta/livekit/plugins/meta/log.py

@@ -0,0 +1,3 @@
+import logging
+
+logger = logging.getLogger("livekit.plugins.meta")

livekit-plugins/livekit-plugins-meta/livekit/plugins/meta/models.py

@@ -0,0 +1,86 @@
+from __future__ import annotations
+
+from typing import Literal
+
+from livekit.agents.language import LanguageCode
+
+from .log import logger
+
+MuseModels = Literal["muse-voice-transcribe-1.0"]
+
+MuseEncoding = Literal["PCM_24KHZ", "PCM_16KHZ"]
+"""Signed 16-bit little-endian mono PCM, at 24 kHz or 16 kHz."""
+
+SAMPLE_RATES: dict[str, int] = {"PCM_24KHZ": 24000, "PCM_16KHZ": 16000}
+
+MuseMode = Literal["PUSH_TO_TALK", "ENDPOINTING", "DIARIZATION"]
+"""PUSH_TO_TALK: the caller delimits the turn. ENDPOINTING: the model detects turn
+boundaries, one turn per speech segment. DIARIZATION: adds speaker attribution."""
+
+MusePartialMode = Literal["CUMULATIVE", "DELTA"]
+"""CUMULATIVE: every ``transcript`` event carries the whole current hypothesis and
+replaces the previous one. DELTA: per-chunk text, for the file endpoint's SSE mode."""
+
+# The 25 languages Meta lists as evaluated for muse-voice-transcribe-1.0. Muse takes
+# `languageBias` as English language *names*, not BCP-47 codes, so this is keyed by the
+# base code LanguageCode.language produces ("en-US" and "cmn-Hans-CN" both reduce here).
+SUPPORTED_LANGUAGES: frozenset[str] = frozenset(
+    {
+        "ar",
+        "bn",
+        "de",
+        "en",
+        "es",
+        "fr",
+        "he",
+        "hi",
+        "id",
+        "it",
+        "ja",
+        "kn",
+        "ko",
+        "mr",
+        "ms",
+        "nl",
+        "pl",
+        "pt",
+        "ta",
+        "te",
+        "th",
+        "tl",
+        "tr",
+        "vi",
+        "zh",
+    }
+)
+
+# LanguageCode.to_language_name() covers 23 of the 25 (lowercased); these two it does
+# not: it calls zh "chinese" where Meta's list says "Mandarin Chinese", and it has no
+# name for Filipino, which is Tagalog on Meta's side.
+_LANGUAGE_NAME_OVERRIDES: dict[str, str] = {"zh": "Mandarin Chinese", "fil": "Tagalog"}
+
+
+def to_language_bias(languages: list[str]) -> list[str]:
+    """Map BCP-47 codes onto the English language names Muse's `languageBias` expects.
+
+    Codes Muse does not list are dropped with a warning rather than forwarded: the
+    handshake is rejected wholesale on an unknown bias entry, which would take down
+    recognition entirely instead of merely losing the hint.
+    """
+    names: list[str] = []
+    for lang in languages:
+        code = LanguageCode(lang)
+        base = code.language
+        name = _LANGUAGE_NAME_OVERRIDES.get(base)
+        if name is None and base in SUPPORTED_LANGUAGES:
+            resolved = code.to_language_name()
+            name = resolved.title() if resolved else None
+        if name is None:
+            logger.warning(
+                "language is not supported by Muse Voice Transcribe, dropping from languageBias",
+                extra={"language": lang},
+            )
+            continue
+        if name not in names:
+            names.append(name)
+    return names