feat(fishaudio): use websocket API for faster inference

livekit-agents/livekit/agents/tts/tts.py

@@ -1192,11 +1192,17 @@ async def _decode_task() -> None:
                         audio_decoder.push(data)
                     elif decode_atask:
                         if isinstance(data, AudioEmitter._FlushSegment):
-                            if audio_decoder:
-                                audio_decoder.end_input()
-                                await decode_atask
-                                _flush_frame()
-                                audio_decoder = None
+                            # don't tear the decoder down here. flush_if_delayed
+                            # can fire mid-stream while a stateful codec
+                            # (WAV/OGG/MP3) is in the middle of a file; ending
+                            # input would discard the parser, and the next
+                            # bytes — a pure PCM/Opus packet continuation
+                            # without a fresh container header — would fail to
+                            # parse against a freshly-created decoder. The only
+                            # purpose of FlushSegment here is to release the
+                            # held-back tail so a slow upstream doesn't starve
+                            # the consumer.
+                            _flush_frame()
 
                         elif isinstance(data, AudioEmitter._EndSegment) and segment_ctx:
                             if audio_decoder:

livekit-agents/livekit/agents/utils/codecs/decoder.py

@@ -461,6 +461,11 @@ def _decode_loop(self) -> None:
                 for f in resampler.resample(None):
                     self._emit_av_frame(f)
 
+        except av.error.EOFError:
+            # Input ended before FFmpeg could parse enough data — typically caused
+            # by early cancellation, a very short utterance, or a premature upstream
+            # close. The container has nothing to emit; this is not an error.
+            logger.debug("audio decoder input ended before any audio could be decoded")
         except Exception:
             logger.exception("error decoding audio")
         finally:

livekit-plugins/livekit-plugins-fishaudio/livekit/plugins/fishaudio/__init__.py

@@ -6,18 +6,16 @@
 - `FISH_API_KEY` for authentication (required)
 """
 
-from fish_audio_sdk.schemas import Backends
-
 from livekit.agents import Plugin
 
 from .log import logger
-from .models import LatencyMode, OutputFormat
+from .models import LatencyMode, OutputFormat, TTSModels
 from .tts import TTS
 from .version import __version__
 
 __all__ = [
     "TTS",
-    "Backends",
+    "TTSModels",
     "OutputFormat",
     "LatencyMode",
     "__version__",

livekit-plugins/livekit-plugins-fishaudio/livekit/plugins/fishaudio/models.py

@@ -1,5 +1,7 @@
 from typing import Literal
 
-OutputFormat = Literal["wav", "pcm", "mp3"]
+TTSModels = Literal["s1", "s2-pro"]
 
-LatencyMode = Literal["normal", "balanced"]
+OutputFormat = Literal["wav", "pcm", "mp3", "opus"]
+
+LatencyMode = Literal["normal", "balanced", "low"]