feat: Add OpenAI-compatible TTS endpoint and engines

- Implements POST /v1/audio/speech endpoint (OpenAI API compatible).
- Integrates Kokoro and XTTS engines (including dependencies and implementations).
- Updates main application to register new engines and router.
- Adds unit tests for OpenAI compatibility.
- Updates requirements.txt for new engines.
This commit is contained in:
2025-12-09 12:45:17 +01:00
parent c06fd677dc
commit fff0252d52
9 changed files with 985 additions and 3 deletions

338
app/engines/kokoro.py Normal file
View File

@ -0,0 +1,338 @@
"""
NovaAi – TTS-Engine-Hub
engines/kokoro.py
Version: v0.1.0
Description:
Kokoro TTS engine adapter: 82M parameter high-quality TTS model.
Synthesizes 24kHz audio using Kokoro library, converts to OGG/MP3 via ffmpeg if needed.
Supports 54 voices across 8 languages with GPU acceleration.
Author: Claude Code (Anthropic)
Date: 2025-12-05
"""
import asyncio
import subprocess
import tempfile
import os
import shutil
import logging
from typing import Optional, List
from .engine_base import TTSEngineBase
from app.config import settings
import ffmpeg
logger = logging.getLogger(__name__)
# Language code mapping for Kokoro
KOKORO_LANG_CODES = {
"kokoro-en-us": "a", # American English
"kokoro-en-gb": "b", # British English
"kokoro-fr": "fr", # French
"kokoro-es": "es", # Spanish
"kokoro-ja": "ja", # Japanese
"kokoro-zh": "zh", # Chinese
"kokoro-it": "it", # Italian
"kokoro-pt": "pt", # Portuguese
"kokoro-hi": "hi", # Hindi
"kokoro-ko": "ko", # Korean
}
# Import voice metadata
from .kokoro_voices import ALL_VOICES, get_voices_for_model, get_voice_info
class KokoroEngine(TTSEngineBase):
def __init__(self):
self.kokoro_available = False
self.pipeline = None
self.current_lang = None
self.ffmpeg_executable = shutil.which("ffmpeg")
self.device = getattr(settings, "KOKORO_DEVICE", "cuda")
self.timeout = getattr(settings, "KOKORO_TIMEOUT_SECONDS", 30)
# Try to import and initialize Kokoro
try:
from kokoro import KPipeline
self.KPipeline = KPipeline
self.kokoro_available = True
logger.info("Kokoro TTS library loaded successfully")
except ImportError as e:
logger.warning(f"Kokoro TTS library not available: {e}")
self.kokoro_available = False
def _get_pipeline(self, lang_code: str):
"""Get or create pipeline for specific language."""
if not self.kokoro_available:
raise RuntimeError("Kokoro library not installed. Install with: pip install kokoro>=0.9.2")
# Reuse pipeline if same language
if self.pipeline is not None and self.current_lang == lang_code:
return self.pipeline
# Create new pipeline for language
try:
logger.info(f"Loading Kokoro pipeline for language code: {lang_code}")
self.pipeline = self.KPipeline(lang_code=lang_code)
self.current_lang = lang_code
return self.pipeline
except Exception as e:
logger.error(f"Failed to load Kokoro pipeline: {e}")
raise RuntimeError(f"Failed to load Kokoro pipeline for {lang_code}: {e}")
def _run_ffmpeg_blocking(self, input_path: str, output_path: str):
"""
Wrapper for blocking ffmpeg call with error capture.
Reused from Piper engine implementation.
"""
try:
stdout, stderr = (
ffmpeg
.input(input_path)
.output(output_path)
.run(overwrite_output=True, capture_stdout=True, capture_stderr=True)
)
if stderr:
logger.debug(f"FFmpeg output: {stderr.decode('utf-8', errors='replace')}")
except ffmpeg.Error as e:
stderr_output = e.stderr.decode('utf-8', errors='replace') if e.stderr else "No error output"
logger.error(f"FFmpeg conversion failed: {input_path} -> {output_path}. Error: {stderr_output}")
raise RuntimeError(
f"FFmpeg conversion failed: {input_path} -> {output_path}. "
f"Error: {stderr_output}"
)
async def synthesize(self, text: str, speaker: str = None, model: str = None, fmt: str = "ogg") -> str:
"""
Synthesize speech from text using Kokoro TTS.
Applies all bug fixes from Piper engine:
- Timeout protection
- Comprehensive temp file cleanup
- FFmpeg error capture
- Enhanced logging
"""
# Validation
if not self.kokoro_available:
raise RuntimeError("Kokoro library not installed. Install with: pip install kokoro>=0.9.2 soundfile")
if not model:
model = "kokoro-en-us" # Default to American English
if model not in KOKORO_LANG_CODES:
raise ValueError(
f"Model '{model}' not supported. Available models: {list(KOKORO_LANG_CODES.keys())}"
)
if not speaker:
speaker = "af_bella" # Default voice
if speaker not in ALL_VOICES:
logger.warning(
f"Voice '{speaker}' not in known voice list. Attempting anyway. "
f"Known voices: {ALL_VOICES[:10]}..."
)
# Get language code
lang_code = KOKORO_LANG_CODES[model]
# Track temp files for cleanup
temp_files_to_cleanup = []
try:
# Get pipeline for language
pipeline = await asyncio.to_thread(self._get_pipeline, lang_code)
# Create WAV temp file
fd, output_wav_path = tempfile.mkstemp(suffix=".wav", prefix="kokoro_")
os.close(fd)
temp_files_to_cleanup.append(output_wav_path)
# Log synthesis details
text_preview = text[:100] + "..." if len(text) > 100 else text
logger.debug(f"Kokoro synthesis: model={model}, voice={speaker}, text_len={len(text)}")
logger.debug(f"Text preview: {text_preview}")
# Generate audio with timeout
try:
audio_data = await asyncio.wait_for(
asyncio.to_thread(self._synthesize_audio, pipeline, text, speaker),
timeout=self.timeout
)
except asyncio.TimeoutError:
logger.error(
f"Kokoro synthesis timed out after {self.timeout}s. "
f"Model: {model}, Voice: {speaker}, Text length: {len(text)}"
)
raise RuntimeError(
f"Kokoro synthesis timed out after {self.timeout}s. "
f"Text length: {len(text)} chars"
)
# Save audio to WAV file
import soundfile as sf
await asyncio.to_thread(sf.write, output_wav_path, audio_data, 24000)
# Verify output created
if not os.path.exists(output_wav_path) or os.path.getsize(output_wav_path) == 0:
raise RuntimeError("Kokoro synthesis failed: output file not created or empty")
logger.info(
f"Kokoro synthesis succeeded: {len(text)} chars -> "
f"{os.path.getsize(output_wav_path)} bytes. Model: {model}, Voice: {speaker}"
)
# Return WAV if requested
fmt = (fmt or "ogg").lower()
if fmt == "wav":
temp_files_to_cleanup.remove(output_wav_path)
return output_wav_path
# FFmpeg conversion
if not self.ffmpeg_executable:
raise RuntimeError("ffmpeg not found, cannot convert audio format.")
# Create converted file temp path
fd_conv, output_other_path = tempfile.mkstemp(suffix=f'.{fmt}', prefix="kokoro_conv_")
os.close(fd_conv)
temp_files_to_cleanup.append(output_other_path)
logger.debug(f"Converting WAV to {fmt}: {output_wav_path} -> {output_other_path}")
# Convert with timeout
try:
await asyncio.wait_for(
asyncio.to_thread(self._run_ffmpeg_blocking, output_wav_path, output_other_path),
timeout=60 # FFmpeg timeout
)
except asyncio.TimeoutError:
logger.error(
f"FFmpeg conversion timed out after 60s. "
f"Input size: {os.path.getsize(output_wav_path)} bytes"
)
raise RuntimeError(
f"FFmpeg conversion timed out after 60s. "
f"Input size: {os.path.getsize(output_wav_path)} bytes"
)
# Verify conversion succeeded
if not os.path.exists(output_other_path) or os.path.getsize(output_other_path) == 0:
raise RuntimeError("FFmpeg conversion failed: output file not created or empty")
logger.info(
f"FFmpeg conversion succeeded: {os.path.getsize(output_wav_path)} bytes (WAV) -> "
f"{os.path.getsize(output_other_path)} bytes ({fmt})"
)
# Success! Remove converted file from cleanup (we're returning it)
temp_files_to_cleanup.remove(output_other_path)
return output_other_path
finally:
# Cleanup all temp files
for temp_file in temp_files_to_cleanup:
try:
if os.path.exists(temp_file):
os.remove(temp_file)
logger.debug(f"Cleaned up temp file: {temp_file}")
except Exception as e:
logger.warning(f"Failed to cleanup temp file {temp_file}: {e}")
def _synthesize_audio(self, pipeline, text: str, voice: str):
"""
Blocking synthesis function (runs in thread).
Generates audio using Kokoro pipeline.
"""
import numpy as np
# Generate audio using pipeline
generator = pipeline(text, voice=voice)
# Collect audio chunks
audio_chunks = []
for gs, ps, audio in generator:
audio_chunks.append(audio)
# Concatenate all chunks
if not audio_chunks:
raise RuntimeError("Kokoro generated no audio chunks")
full_audio = np.concatenate(audio_chunks)
return full_audio
def list_models(self) -> List[str]:
"""Return available Kokoro language models."""
return list(KOKORO_LANG_CODES.keys())
def list_voices(self, model: str = None) -> List[str]:
"""Return available Kokoro voices, optionally filtered by model/language."""
if model and model in KOKORO_LANG_CODES:
# Return voices for specific language
return sorted(get_voices_for_model(model))
else:
# Return all voices
return sorted(ALL_VOICES)
def healthcheck(self):
"""Return health/status info for Kokoro engine."""
status = "ok" if self.kokoro_available else "not_available"
details = {
"status": status,
"engine": "kokoro",
"library_available": self.kokoro_available,
"device": self.device if self.kokoro_available else None,
}
if not self.kokoro_available:
details["error"] = "Kokoro library not installed. Install with: pip install kokoro>=0.9.2 soundfile"
return details
async def selftest(self):
"""Run self-test to verify Kokoro is working."""
if not self.kokoro_available:
return {
"selftest": False,
"error": "Kokoro library not installed",
"engine": "kokoro"
}
try:
# Test synthesis with default model and voice
test_text = "This is a Kokoro selftest."
audio_file = await self.synthesize(
test_text,
speaker="af_bella",
model="kokoro-en-us",
fmt="wav"
)
selftest_passed = os.path.exists(audio_file) and os.path.getsize(audio_file) > 0
if selftest_passed:
os.remove(audio_file)
return {
"selftest": selftest_passed,
"models": self.list_models(),
"voices_count": len(self.list_voices()),
"engine": "kokoro"
}
except Exception as e:
return {
"selftest": False,
"error": str(e),
"engine": "kokoro"
}
if __name__ == "__main__":
async def main():
engine = KokoroEngine()
print("Healthcheck:", engine.healthcheck())
print("Models:", engine.list_models())
print("Voices:", engine.list_voices()[:10], "...")
print("Selftest:", await engine.selftest())
asyncio.run(main())