[Bugfix] Accept 'speaker' as alias for 'voice' in TTS speech API (#2424)

Signed-off-by: marksverdhei <marksverdhei@hotmail.com>
Signed-off-by: marksverdhai <249650165+marksverdhai@users.noreply.github.com>
Co-authored-by: marksverdhai <249650165+marksverdhai@users.noreply.github.com>
This commit is contained in:
Markus / Markandmarksverdhai authored and GitHub committed 2026-04-07 15:07:01 +00:00
1 parent 408365fa0d
commit feefdaee9f
4 files changed
+255 -16

No files matched your search

@@ -2,7 +2,7 @@ import math
from typing import Literal
import numpy as np
from pydantic import BaseModel, Field, field_validator, model_validator
from pydantic import AliasChoices, BaseModel, Field, field_validator, model_validator
_MAX_EMBEDDING_DIM = 8192
@@ -10,8 +10,12 @@ _MAX_EMBEDDING_DIM = 8192
class OpenAICreateSpeechRequest(BaseModel):
input: str
model: str | None = None
# Accept both "voice" (OpenAI convention) and "speaker" (model/internal
# convention) as input keys. Intentionally global — all TTS backends
# (Qwen3-TTS, Voxtral, Fish Speech) use this field for the speaker name.
voice: str | None = Field(
default=None,
validation_alias=AliasChoices("voice", "speaker"),
description="Speaker/voice to use. For Qwen3-TTS: vivian, ryan, aiden, etc.",
)
instructions: str | None = Field(
+74 -14
View File
@@ -467,6 +467,48 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
logger.error(f"Could not read audio file for voice {voice_name}: {e}")
return None
def _get_uploaded_speaker_embedding(self, voice_name: str) -> list[float] | None:
"""Load pre-computed speaker embedding for an uploaded voice.
Returns the embedding as a list of floats, or None if the voice
was not uploaded with an embedding (i.e. it has audio instead).
"""
voice_name_lower = voice_name.lower()
if voice_name_lower not in self.uploaded_speakers:
return None
speaker_info = self.uploaded_speakers[voice_name_lower]
if speaker_info.get("embedding_source") != "direct":
return None
cache_file = speaker_info.get("cache_file")
if not cache_file or not Path(cache_file).exists():
logger.warning("Embedding file not found for voice %s: %s", voice_name, cache_file)
return None
if not _validate_path_within_directory(Path(cache_file), self.uploaded_speakers_dir):
logger.error("Cache file path traversal detected for voice %s: %s", voice_name, cache_file)
return None
try:
from safetensors.torch import load_file
except ImportError:
logger.error(
"The 'safetensors' package is required to load speaker embeddings. "
"Install it with: pip install safetensors"
)
return None
try:
tensors = load_file(cache_file)
if "speaker_embedding" not in tensors:
logger.warning("Key 'speaker_embedding' not found in %s for voice %s", cache_file, voice_name)
return None
return tensors["speaker_embedding"].squeeze().tolist()
except Exception as e:
logger.error("Could not load embedding for voice %s: %s", voice_name, e)
return None
async def upload_voice(
self,
audio_file: UploadFile,
@@ -858,11 +900,17 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
# voice is not None
voice_lower = request.voice.lower()
if voice_lower in self.uploaded_speakers:
# Check if audio file exists for uploaded speaker
# Check if data file exists for uploaded speaker
speaker_info = self.uploaded_speakers[voice_lower]
file_path = Path(speaker_info["file_path"])
if not file_path.exists():
return f"Audio file for uploaded speaker '{request.voice}' not found on disk"
return f"Data file for uploaded speaker '{request.voice}' not found on disk"
# For embedding-uploaded voices, verify the cache is ready
if speaker_info.get("embedding_source") == "direct":
cache_file = speaker_info.get("cache_file")
if not cache_file or not Path(cache_file).exists():
status = speaker_info.get("cache_status", "unknown")
return f"Speaker embedding for '{request.voice}' is not yet ready (cache_status='{status}')"
else:
# need ref_audio for built-in speaker
if request.ref_audio is None:
@@ -1107,20 +1155,32 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
# Uploaded voices use task_type="Base" (CustomVoice requires built-in spk_id).
# If ref_text was provided at upload time, use in-context cloning; otherwise x_vector only.
if request.voice.lower() in self.uploaded_speakers and request.ref_audio is None:
audio_data = self._get_uploaded_audio_data(request.voice)
if not audio_data:
raise ValueError(f"Audio file for uploaded voice '{request.voice}' is missing or corrupted")
speaker_info = self.uploaded_speakers[request.voice.lower()]
stored_ref_text = speaker_info.get("ref_text")
params["ref_audio"] = [audio_data]
params["task_type"] = ["Base"]
params["voice_created_at"] = [speaker_info.get("created_at", 0)]
if stored_ref_text:
params["ref_text"] = [stored_ref_text]
params["x_vector_only_mode"] = [False]
# Check if this voice was uploaded with a pre-computed embedding.
# Populate request.speaker_embedding so the existing code path
# (below) handles voice_clone_prompt and x_vector_only_mode.
embedding = self._get_uploaded_speaker_embedding(request.voice)
if embedding is not None:
request.speaker_embedding = embedding
params["task_type"] = ["Base"]
logger.info("Auto-set speaker_embedding for uploaded voice: %s", request.voice)
else:
params["x_vector_only_mode"] = [True]
logger.info("Auto-set ref_audio for uploaded voice: %s (icl=%s)", request.voice, bool(stored_ref_text))
audio_data = self._get_uploaded_audio_data(request.voice)
if not audio_data:
raise ValueError(f"Audio file for uploaded voice '{request.voice}' is missing or corrupted")
stored_ref_text = speaker_info.get("ref_text")
params["ref_audio"] = [audio_data]
params["task_type"] = ["Base"]
params["voice_created_at"] = [speaker_info.get("created_at", 0)]
if stored_ref_text:
params["ref_text"] = [stored_ref_text]
params["x_vector_only_mode"] = [False]
else:
params["x_vector_only_mode"] = [True]
logger.info(
"Auto-set ref_audio for uploaded voice: %s (icl=%s)", request.voice, bool(stored_ref_text)
)
elif params["task_type"][0] == "CustomVoice":
params["speaker"] = ["Vivian"] # Default for CustomVoice