livekit-plugins-elevenlabs 0.4.dev2__tar.gz → 0.5.dev0__tar.gz

{livekit_plugins_elevenlabs-0.4.dev2 → livekit_plugins_elevenlabs-0.5.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: livekit-plugins-elevenlabs
-Version: 0.4.dev2
+Version: 0.5.dev0
 Summary: Agent Framework plugin for voice synthesis with ElevenLabs' API.
 Home-page: https://github.com/livekit/agents
 License: Apache-2.0
@@ -20,7 +20,7 @@ Classifier: Programming Language :: Python :: 3 :: Only
 Requires-Python: >=3.9.0
 Description-Content-Type: text/markdown
 Requires-Dist: livekit~=0.11
-Requires-Dist: livekit-agents~=0.6.dev1
+Requires-Dist: livekit-agents~=0.7.dev0
 Requires-Dist: aiohttp>=3.8.5
 
 # LiveKit Plugins Elevenlabs

livekit_plugins_elevenlabs-0.5.dev0/livekit/plugins/elevenlabs/tts.py

@@ -0,0 +1,487 @@
+# Copyright 2023 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.
+
+from __future__ import annotations
+
+import asyncio
+import base64
+import contextlib
+import dataclasses
+import json
+import os
+from dataclasses import dataclass
+from typing import List, Optional
+
+import aiohttp
+from livekit import rtc
+from livekit.agents import aio, tokenize, tts, utils
+
+from .log import logger
+from .models import TTSModels
+
+
+@dataclass
+class VoiceSettings:
+    stability: float  # [0.0 - 1.0]
+    similarity_boost: float  # [0.0 - 1.0]
+    style: float | None = None  # [0.0 - 1.0]
+    use_speaker_boost: bool | None = False
+
+
+@dataclass
+class Voice:
+    id: str
+    name: str
+    category: str
+    settings: VoiceSettings | None = None
+
+
+DEFAULT_VOICE = Voice(
+    id="EXAVITQu4vr4xnSDxMaL",
+    name="Bella",
+    category="premade",
+    settings=VoiceSettings(
+        stability=0.71, similarity_boost=0.5, style=0.0, use_speaker_boost=True
+    ),
+)
+
+API_BASE_URL_V1 = "https://api.elevenlabs.io/v1"
+AUTHORIZATION_HEADER = "xi-api-key"
+
+
+@dataclass
+class _TTSOptions:
+    api_key: str
+    voice: Voice
+    model_id: TTSModels
+    base_url: str
+    sample_rate: int
+    streaming_latency: int
+    word_tokenizer: tokenize.WordTokenizer
+    chunk_length_schedule: list[int]
+
+
+class TTS(tts.TTS):
+    def __init__(
+        self,
+        *,
+        voice: Voice = DEFAULT_VOICE,
+        model_id: TTSModels = "eleven_turbo_v2",
+        api_key: str | None = None,
+        base_url: str | None = None,
+        sample_rate: int = 24000,
+        streaming_latency: int = 3,
+        word_tokenizer: tokenize.WordTokenizer = tokenize.basic.WordTokenizer(
+            ignore_punctuation=False  # punctuation can help for intonation
+        ),
+        # default value of 11labs is [120, 160, 250, 290], but we want faster responses by default
+        # (range is 50-500)
+        chunk_length_schedule: list[int] = [80, 120, 200, 260],
+        http_session: aiohttp.ClientSession | None = None,
+    ) -> None:
+        super().__init__(
+            streaming_supported=True, sample_rate=sample_rate, num_channels=1
+        )
+        api_key = api_key or os.environ.get("ELEVEN_API_KEY")
+        if not api_key:
+            raise ValueError("ELEVEN_API_KEY must be set")
+
+        self._opts = _TTSOptions(
+            voice=voice,
+            model_id=model_id,
+            api_key=api_key,
+            base_url=base_url or API_BASE_URL_V1,
+            sample_rate=sample_rate,
+            streaming_latency=streaming_latency,
+            word_tokenizer=word_tokenizer,
+            chunk_length_schedule=chunk_length_schedule,
+        )
+        self._session = http_session
+
+    def _ensure_session(self) -> aiohttp.ClientSession:
+        if not self._session:
+            self._session = utils.http_session()
+
+        return self._session
+
+    async def list_voices(self) -> List[Voice]:
+        async with self._ensure_session().get(
+            f"{self._opts.base_url}/voices",
+            headers={AUTHORIZATION_HEADER: self._opts.api_key},
+        ) as resp:
+            return _dict_to_voices_list(await resp.json())
+
+    def synthesize(
+        self,
+        text: str,
+    ) -> "ChunkedStream":
+        return ChunkedStream(text, self._opts, self._ensure_session())
+
+    def stream(
+        self,
+    ) -> "SynthesizeStream":
+        return SynthesizeStream(self._ensure_session(), self._opts)
+
+
+class ChunkedStream(tts.ChunkedStream):
+    """Synthesize using the chunked api endpoint"""
+
+    def __init__(
+        self, text: str, opts: _TTSOptions, session: aiohttp.ClientSession
+    ) -> None:
+        self._opts = opts
+        self._text = text
+        self._session = session
+        self._task: asyncio.Task | None = None
+        self._queue = asyncio.Queue[Optional[tts.SynthesizedAudio]]()
+
+    def _synthesize_url(self) -> str:
+        base_url = self._opts.base_url
+        voice_id = self._opts.voice.id
+        model_id = self._opts.model_id
+        sample_rate = self._opts.sample_rate
+        latency = self._opts.streaming_latency
+        url = (
+            f"{base_url}/text-to-speech/{voice_id}/stream?"
+            f"model_id={model_id}&output_format=pcm_{sample_rate}&optimize_streaming_latency={latency}"
+        )
+        return url
+
+    async def _main_task(self):
+        try:
+            await self._run()
+        except Exception:
+            logger.exception("11labs main task failed in chunked stream")
+        finally:
+            self._queue.put_nowait(None)
+
+    async def _run(self) -> None:
+        async with self._session.post(
+            self._synthesize_url(),
+            headers={AUTHORIZATION_HEADER: self._opts.api_key},
+            json=dict(
+                text=self._text,
+                model_id=self._opts.model_id,
+                voice_settings=(
+                    dataclasses.asdict(self._opts.voice.settings)
+                    if self._opts.voice.settings
+                    else None
+                ),
+            ),
+        ) as resp:
+            # avoid very small frames. chunk by 10ms 16bits
+            bytes_per_frame = (self._opts.sample_rate // 100) * 2
+            buf = bytearray()
+            async for data, _ in resp.content.iter_chunks():
+                buf.extend(data)
+
+                while len(buf) >= bytes_per_frame:
+                    frame_data = buf[:bytes_per_frame]
+                    buf = buf[bytes_per_frame:]
+
+                    self._queue.put_nowait(
+                        tts.SynthesizedAudio(
+                            text=self._text,
+                            data=rtc.AudioFrame(
+                                data=frame_data,
+                                sample_rate=self._opts.sample_rate,
+                                num_channels=1,
+                                samples_per_channel=len(frame_data) // 2,
+                            ),
+                        )
+                    )
+
+            # send any remaining data
+            if len(buf) > 0:
+                self._queue.put_nowait(
+                    tts.SynthesizedAudio(
+                        text=self._text,
+                        data=rtc.AudioFrame(
+                            data=buf,
+                            sample_rate=self._opts.sample_rate,
+                            num_channels=1,
+                            samples_per_channel=len(buf) // 2,
+                        ),
+                    )
+                )
+
+    async def __anext__(self) -> tts.SynthesizedAudio:
+        if not self._task:
+            self._task = asyncio.create_task(self._main_task())
+
+        frame = await self._queue.get()
+        if frame is None:
+            raise StopAsyncIteration
+
+        return frame
+
+    async def aclose(self) -> None:
+        if not self._task:
+            return
+
+        self._task.cancel()
+        with contextlib.suppress(asyncio.CancelledError):
+            await self._task
+
+
+class SynthesizeStream(tts.SynthesizeStream):
+    """Streamed API using websockets"""
+
+    @dataclass
+    class _SegmentConnection:
+        audio_rx: aio.ChanReceiver[tts.SynthesizedAudio]
+        task: asyncio.Task
+
+    def __init__(
+        self,
+        session: aiohttp.ClientSession,
+        opts: _TTSOptions,
+        max_retry_per_segment: int = 3,
+    ):
+        self._opts = opts
+        self._session = session
+        self._main_task = asyncio.create_task(self._run(max_retry_per_segment))
+        self._event_queue = asyncio.Queue[Optional[tts.SynthesisEvent]]()
+        self._closed = False
+        self._word_stream = opts.word_tokenizer.stream()
+
+    def _stream_url(self) -> str:
+        base_url = self._opts.base_url
+        voice_id = self._opts.voice.id
+        model_id = self._opts.model_id
+        sample_rate = self._opts.sample_rate
+        latency = self._opts.streaming_latency
+        url = (
+            f"{base_url}/text-to-speech/{voice_id}/stream-input?"
+            f"model_id={model_id}&output_format=pcm_{sample_rate}&optimize_streaming_latency={latency}"
+        )
+
+        return url
+
+    def push_text(self, token: str | None) -> None:
+        if self._closed:
+            raise ValueError("cannot push to a closed stream")
+
+        if token is None:
+            self._word_stream.mark_segment_end()
+            return
+
+        self._word_stream.push_text(token)
+
+    async def aclose(self, *, wait: bool = True) -> None:
+        self._closed = True
+        await self._word_stream.aclose()
+
+        if not wait:
+            self._main_task.cancel()
+
+        with contextlib.suppress(asyncio.CancelledError):
+            await self._main_task
+
+    async def _run(self, max_retry_per_segment: int) -> None:
+        conns_q = asyncio.Queue[Optional[SynthesizeStream._SegmentConnection]]()
+
+        async def _forward_events() -> None:
+            """forward events from the ws connections to the event queue.
+            This is used to keep the right order."""
+            while True:
+                c = await conns_q.get()
+                if c is None:
+                    break  # no more segment, stream closed
+
+                self._event_queue.put_nowait(
+                    tts.SynthesisEvent(type=tts.SynthesisEventType.STARTED)
+                )
+
+                async for frame in c.audio_rx:
+                    self._event_queue.put_nowait(
+                        tts.SynthesisEvent(
+                            type=tts.SynthesisEventType.AUDIO, audio=frame
+                        )
+                    )
+
+                self._event_queue.put_nowait(
+                    tts.SynthesisEvent(type=tts.SynthesisEventType.FINISHED)
+                )
+
+        async def _read_tokens() -> None:
+            """read tokens from the word stream and create connections for each segment,
+            (this also allows concurrent connections to 11labs)"""
+
+            cur_segment: SynthesizeStream._SegmentConnection | None = None
+            token_tx: aio.ChanSender[str] | None = None
+            async for ev in self._word_stream:
+                if ev.type == tokenize.TokenEventType.STARTED:
+                    token_tx, token_rx = aio.channel()
+                    audio_tx: aio.ChanSender[tts.SynthesizedAudio]
+                    audio_rx: aio.ChanReceiver[tts.SynthesizedAudio]
+                    audio_tx, audio_rx = aio.channel()
+                    task = asyncio.create_task(
+                        self._run_ws(max_retry_per_segment, audio_tx, token_rx)
+                    )
+                    cur_segment = SynthesizeStream._SegmentConnection(audio_rx, task)
+                    conns_q.put_nowait(cur_segment)
+                elif ev.type == tokenize.TokenEventType.TOKEN:
+                    assert token_tx is not None
+                    token_tx.send_nowait(ev.token)
+                elif ev.type == tokenize.TokenEventType.FINISHED:
+                    assert token_tx is not None
+                    token_tx.close()
+                    cur_segment = token_tx = None
+
+            conns_q.put_nowait(None)
+
+        try:
+            await asyncio.gather(_forward_events(), _read_tokens())
+        except Exception:
+            logger.exception("11labs task failed")
+
+        self._event_queue.put_nowait(None)
+
+    async def _run_ws(
+        self,
+        max_retry: int,
+        audio_tx: aio.ChanSender[tts.SynthesizedAudio],
+        token_rx: aio.ChanReceiver[str],
+    ) -> None:
+        # try to connect to 11labs
+        ws_conn: aiohttp.ClientWebSocketResponse | None = None
+        for try_i in range(max_retry):
+            try:
+                ws_conn = await self._session.ws_connect(
+                    self._stream_url(),
+                    headers={AUTHORIZATION_HEADER: self._opts.api_key},
+                )
+
+                voice_settings = None
+                if self._opts.voice.settings is not None:
+                    voice_settings = dataclasses.asdict(self._opts.voice.settings)
+
+                init_pkt = dict(
+                    text=" ",
+                    try_trigger_generation=True,
+                    voice_settings=voice_settings,
+                    generation_config=dict(
+                        chunk_length_schedule=self._opts.chunk_length_schedule,
+                    ),
+                )
+                await ws_conn.send_str(json.dumps(init_pkt))
+            except Exception:
+                if try_i + 1 == max_retry:
+                    logger.exception(
+                        f"failed to connect to 11labs after {max_retry} retries"
+                    )
+                    return
+
+                retry_delay = min(try_i * 5, 5)  # max 5s
+                logger.warning(
+                    f"failed to connect to 11labs, retrying in {retry_delay}s"
+                )
+                await asyncio.sleep(retry_delay)
+
+        assert ws_conn is not None
+
+        all_tokens_consumed = False
+
+        async def send_task():
+            async for token in token_rx:
+                if token == "":
+                    continue  # empty token is closing the stream in 11labs protocol
+
+                # try_trigger_generation=True is a bad practice, we expose
+                # chunk_length_schedule instead
+                data_pkt = dict(
+                    text=f"{token} ",  # must always end with a space
+                    try_trigger_generation=False,
+                )
+                await ws_conn.send_str(json.dumps(data_pkt))
+
+            # no more token, mark eos
+            flush_pkt = dict(
+                text="",
+            )
+            await ws_conn.send_str(json.dumps(flush_pkt))
+
+            nonlocal all_tokens_consumed
+            all_tokens_consumed = True
+
+        async def recv_task():
+            while True:
+                msg = await ws_conn.receive()
+                if msg.type in (
+                    aiohttp.WSMsgType.CLOSED,
+                    aiohttp.WSMsgType.CLOSE,
+                    aiohttp.WSMsgType.CLOSING,
+                ):
+                    if all_tokens_consumed:
+                        return  # close is expected
+
+                    raise Exception(
+                        "11labs connection closed unexpectedly, not all tokens have been consumed"
+                    )
+
+                if msg.type != aiohttp.WSMsgType.TEXT:
+                    # audio frames are serialized in base64..
+                    logger.warning("unexpected 11labs message type %s", msg.type)
+                    continue
+
+                data: dict = json.loads(msg.data)
+                if data.get("audio"):
+                    b64data = base64.b64decode(data["audio"])
+
+                    frame = rtc.AudioFrame(
+                        data=b64data,
+                        sample_rate=self._opts.sample_rate,
+                        num_channels=1,
+                        samples_per_channel=len(b64data) // 2,
+                    )
+
+                    text = ""
+                    if data.get("alignment"):
+                        text = data["alignment"].get("chars", "")
+
+                    audio_tx.send_nowait(tts.SynthesizedAudio(text=text, data=frame))
+                    continue
+                elif data.get("isFinal"):
+                    return  # last message
+
+                logger.error("unexpected 11labs message %s", data)
+
+        try:
+            await asyncio.gather(send_task(), recv_task())
+        except Exception:
+            logger.exception("11labs ws connection failed")
+        finally:
+            audio_tx.close()
+
+    async def __anext__(self) -> tts.SynthesisEvent:
+        evt = await self._event_queue.get()
+        if evt is None:
+            raise StopAsyncIteration
+
+        return evt
+
+
+def _dict_to_voices_list(data: dict) -> List[Voice]:
+    voices = []
+    for voice in data["voices"]:
+        voices.append(
+            Voice(
+                id=voice["voice_id"],
+                name=voice["name"],
+                category=voice["category"],
+                settings=None,
+            )
+        )
+    return voices

{livekit_plugins_elevenlabs-0.4.dev2 → livekit_plugins_elevenlabs-0.5.dev0}/livekit/plugins/elevenlabs/version.py

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "0.4.dev2"
+__version__ = "0.5.dev0"

{livekit_plugins_elevenlabs-0.4.dev2 → livekit_plugins_elevenlabs-0.5.dev0}/livekit_plugins_elevenlabs.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: livekit-plugins-elevenlabs
-Version: 0.4.dev2
+Version: 0.5.dev0
 Summary: Agent Framework plugin for voice synthesis with ElevenLabs' API.
 Home-page: https://github.com/livekit/agents
 License: Apache-2.0
@@ -20,7 +20,7 @@ Classifier: Programming Language :: Python :: 3 :: Only
 Requires-Python: >=3.9.0
 Description-Content-Type: text/markdown
 Requires-Dist: livekit~=0.11
-Requires-Dist: livekit-agents~=0.6.dev1
+Requires-Dist: livekit-agents~=0.7.dev0
 Requires-Dist: aiohttp>=3.8.5
 
 # LiveKit Plugins Elevenlabs

{livekit_plugins_elevenlabs-0.4.dev2 → livekit_plugins_elevenlabs-0.5.dev0}/livekit_plugins_elevenlabs.egg-info/requires.txt

@@ -1,3 +1,3 @@
 livekit~=0.11
-livekit-agents~=0.6.dev1
+livekit-agents~=0.7.dev0
 aiohttp>=3.8.5

{livekit_plugins_elevenlabs-0.4.dev2 → livekit_plugins_elevenlabs-0.5.dev0}/setup.py

@@ -51,7 +51,7 @@ setuptools.setup(
     python_requires=">=3.9.0",
     install_requires=[
         "livekit ~= 0.11",
-        "livekit-agents~=0.6.dev1",
+        "livekit-agents~=0.7.dev0",
         "aiohttp >= 3.8.5",
     ],
     package_data={

livekit_plugins_elevenlabs-0.4.dev2/livekit/plugins/elevenlabs/tts.py

@@ -1,392 +0,0 @@
-# Copyright 2023 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.
-
-import asyncio
-import base64
-import contextlib
-import dataclasses
-import json
-import os
-from dataclasses import dataclass
-from typing import AsyncIterable, List
-
-import aiohttp
-from livekit import rtc
-from livekit.agents import aio, tts
-
-from .log import logger
-from .models import TTSModels
-
-
-@dataclass
-class VoiceSettings:
-    stability: float  # [0.0 - 1.0]
-    similarity_boost: float  # [0.0 - 1.0]
-    style: float | None = None  # [0.0 - 1.0]
-    use_speaker_boost: bool | None = False
-
-
-@dataclass
-class Voice:
-    id: str
-    name: str
-    category: str
-    settings: VoiceSettings | None = None
-
-
-DEFAULT_VOICE = Voice(
-    id="EXAVITQu4vr4xnSDxMaL",
-    name="Bella",
-    category="premade",
-    settings=VoiceSettings(
-        stability=0.71, similarity_boost=0.5, style=0.0, use_speaker_boost=True
-    ),
-)
-
-API_BASE_URL_V1 = "https://api.elevenlabs.io/v1"
-AUTHORIZATION_HEADER = "xi-api-key"
-
-
-@dataclass
-class TTSOptions:
-    api_key: str
-    voice: Voice
-    model_id: TTSModels
-    base_url: str
-    sample_rate: int
-    latency: int
-
-
-class TTS(tts.TTS):
-    def __init__(
-        self,
-        *,
-        voice: Voice = DEFAULT_VOICE,
-        model_id: TTSModels = "eleven_turbo_v2",
-        api_key: str | None = None,
-        base_url: str | None = None,
-        sample_rate: int = 24000,
-        latency: int = 3,
-    ) -> None:
-        super().__init__(
-            streaming_supported=True, sample_rate=sample_rate, num_channels=1
-        )
-        api_key = api_key or os.environ.get("ELEVEN_API_KEY")
-        if not api_key:
-            raise ValueError("ELEVEN_API_KEY must be set")
-
-        self._session = aiohttp.ClientSession()
-        self._opts = TTSOptions(
-            voice=voice,
-            model_id=model_id,
-            api_key=api_key,
-            base_url=base_url or API_BASE_URL_V1,
-            sample_rate=sample_rate,
-            latency=latency,
-        )
-
-    async def list_voices(self) -> List[Voice]:
-        async with self._session.get(
-            f"{self._opts.base_url}/voices",
-            headers={AUTHORIZATION_HEADER: self._opts.api_key},
-        ) as resp:
-            data = await resp.json()
-            return dict_to_voices_list(data)
-
-    def synthesize(
-        self,
-        text: str,
-    ) -> AsyncIterable[tts.SynthesizedAudio]:
-        voice = self._opts.voice
-        url = f"{self._opts.base_url}/text-to-speech/{voice.id}?output_format=pcm_{self._opts.sample_rate}"
-
-        async def generator():
-            try:
-                async with self._session.post(
-                    url,
-                    headers={AUTHORIZATION_HEADER: self._opts.api_key},
-                    json=dict(
-                        text=text,
-                        model_id=self._opts.model_id,
-                        voice_settings=dataclasses.asdict(voice.settings)
-                        if voice.settings
-                        else None,
-                    ),
-                ) as resp:
-                    data = await resp.read()
-                    yield tts.SynthesizedAudio(
-                        text=text,
-                        data=rtc.AudioFrame(
-                            data=data,
-                            sample_rate=self._opts.sample_rate,
-                            num_channels=1,
-                            samples_per_channel=len(data) // 2,  # 16-bit
-                        ),
-                    )
-            except Exception as e:
-                logger.error(f"failed to synthesize: {e}")
-
-        return generator()
-
-    def stream(
-        self,
-    ) -> "SynthesizeStream":
-        return SynthesizeStream(self._session, self._opts)
-
-
-class SynthesizeStream(tts.SynthesizeStream):
-    _STREAM_EOS = ""
-
-    def __init__(
-        self,
-        session: aiohttp.ClientSession,
-        opts: TTSOptions,
-        max_retry: int = 32,
-    ):
-        self._opts = opts
-        self._session = session
-
-        self._queue = asyncio.Queue[str | None]()
-        self._event_queue = asyncio.Queue[tts.SynthesisEvent | None]()
-        self._closed = False
-        self._text = ""
-
-        self._main_task = asyncio.create_task(self._run(max_retry))
-
-    def _stream_url(self) -> str:
-        base_url = self._opts.base_url
-        voice_id = self._opts.voice.id
-        model_id = self._opts.model_id
-        sample_rate = self._opts.sample_rate
-        latency = self._opts.latency
-        return f"{base_url}/text-to-speech/{voice_id}/stream-input?model_id={model_id}&output_format=pcm_{sample_rate}&optimize_streaming_latency={latency}"
-
-    def push_text(self, token: str | None) -> None:
-        if self._closed:
-            raise ValueError("cannot push to a closed stream")
-
-        if token is None:
-            self._flush_if_needed()
-            return
-
-        if len(token) == 0:
-            # 11labs marks the EOS with an empty string, avoid users from pushing empty strings
-            return
-
-        # TODO: Naive word boundary detection may not be good enough for all languages
-        # fmt: off
-        splitters = (".", ",", "?", "!", ";", ":", "—", "-", "(", ")", "[", "]", "}", " ")
-        # fmt: on
-
-        self._text += token
-
-        while True:
-            last_split = -1
-            for i, c in enumerate(self._text):
-                if c in splitters:
-                    last_split = i
-                    break
-
-            if last_split == -1:
-                break
-
-            seg = self._text[: last_split + 1]
-            seg = seg.strip() + " "  # 11labs expects a space at the end
-            self._queue.put_nowait(seg)
-            self._text = self._text[last_split + 1 :]
-
-    async def aclose(self, *, wait: bool = True) -> None:
-        self._flush_if_needed()
-        self._queue.put_nowait(None)
-        self._closed = True
-
-        if not wait:
-            self._main_task.cancel()
-
-        with contextlib.suppress(asyncio.CancelledError):
-            await self._main_task
-
-    def _flush_if_needed(self) -> None:
-        seg = self._text.strip()
-        if len(seg) > 0:
-            self._queue.put_nowait(seg + " ")
-
-        self._text = ""
-        self._queue.put_nowait(SynthesizeStream._STREAM_EOS)
-
-    async def _run(self, max_retry: int) -> None:
-        retry_count = 0
-        ws: aiohttp.ClientWebSocketResponse | None = None
-        ws_task: asyncio.Task | None = None
-        data_tx: aio.ChanSender[str] | None = None
-
-        try:
-            while True:
-                ws_connected = ws is not None and not ws.closed
-                try:
-                    data = await self._queue.get()
-
-                    if data is None:
-                        if ws_task is not None:
-                            await ws_task
-                        break
-
-                    if not ws_connected:
-                        if data == SynthesizeStream._STREAM_EOS:
-                            continue
-
-                        with contextlib.suppress(asyncio.CancelledError):
-                            if ws_task is not None:
-                                await ws_task
-
-                        ws = await self._session.ws_connect(
-                            self._stream_url(),
-                            headers={AUTHORIZATION_HEADER: self._opts.api_key},
-                        )
-                        data_tx, data_rx = aio.channel()
-                        ws_task = asyncio.create_task(self._run_ws(ws, data_rx))
-
-                    assert data_tx is not None
-                    assert ws_task is not None
-                    assert ws is not None
-
-                    data_tx.send_nowait(data)
-
-                except Exception:
-                    if retry_count >= max_retry:
-                        logger.exception(
-                            f"failed to connect to 11labs after {max_retry} retries"
-                        )
-                        break
-
-                    retry_delay = min(retry_count * 5, 5)  # max 5s
-                    retry_count += 1
-
-                    logger.warning(
-                        f"failed to connect to 11labs, retrying in {retry_delay}s"
-                    )
-                    await asyncio.sleep(retry_delay)
-
-        except Exception:
-            logger.exception("11labs task failed")
-        finally:
-            with contextlib.suppress(asyncio.CancelledError):
-                if ws_task is not None:
-                    ws_task.cancel()
-                    await ws_task
-
-            self._event_queue.put_nowait(None)
-
-    async def _run_ws(
-        self, ws: aiohttp.ClientWebSocketResponse, data_rx: aio.ChanReceiver[str]
-    ) -> None:
-        closing_ws = False
-
-        self._event_queue.put_nowait(
-            tts.SynthesisEvent(type=tts.SynthesisEventType.STARTED)
-        )
-
-        async def send_task():
-            nonlocal closing_ws
-
-            # 11labs stream must be initialized with a space
-            voice = self._opts.voice
-            voice_settings = (
-                dataclasses.asdict(voice.settings) if voice.settings else None
-            )
-            init_pkt = dict(
-                text=" ",
-                voice_settings=voice_settings,
-            )
-            await ws.send_str(json.dumps(init_pkt))
-
-            while True:
-                data = await data_rx.recv()
-                data_pkt = dict(
-                    text=data,
-                    try_trigger_generation=True,
-                )
-                if data == SynthesizeStream._STREAM_EOS:
-                    closing_ws = True
-
-                await ws.send_str(json.dumps(data_pkt))
-
-                if closing_ws:
-                    return
-
-        async def recv_task():
-            nonlocal closing_ws
-            while True:
-                msg = await ws.receive()
-                if msg.type in (
-                    aiohttp.WSMsgType.CLOSED,
-                    aiohttp.WSMsgType.CLOSE,
-                    aiohttp.WSMsgType.CLOSING,
-                ):
-                    if closing_ws:  # close is expected
-                        return
-
-                    raise Exception("11labs connection closed unexpectedly")
-
-                if msg.type != aiohttp.WSMsgType.TEXT:
-                    logger.warning("unexpected 11labs message type %s", msg.type)
-                    continue
-
-                data: dict = json.loads(msg.data)
-                if data.get("audio"):
-                    b64data = base64.b64decode(data["audio"])
-                    frame = rtc.AudioFrame(
-                        data=b64data,
-                        sample_rate=self._opts.sample_rate,
-                        num_channels=1,
-                        samples_per_channel=len(b64data) // 2,
-                    )
-                    self._event_queue.put_nowait(
-                        tts.SynthesisEvent(
-                            type=tts.SynthesisEventType.AUDIO,
-                            audio=tts.SynthesizedAudio(text="", data=frame),
-                        )
-                    )
-                elif data.get("isFinal"):
-                    return
-
-        try:
-            await asyncio.gather(send_task(), recv_task())
-        except Exception:
-            logger.exception("11labs connection failed")
-        finally:
-            self._event_queue.put_nowait(
-                tts.SynthesisEvent(type=tts.SynthesisEventType.FINISHED)
-            )
-
-    async def __anext__(self) -> tts.SynthesisEvent:
-        evt = await self._event_queue.get()
-        if evt is None:
-            raise StopAsyncIteration
-
-        return evt
-
-
-def dict_to_voices_list(data: dict) -> List[Voice]:
-    voices = []
-    for voice in data["voices"]:
-        voices.append(
-            Voice(
-                id=voice["voice_id"],
-                name=voice["name"],
-                category=voice["category"],
-                settings=None,
-            )
-        )
-    return voices