""" TTS runtime wrapper (container-scoped singleton). In Modal, each container should load the model once and then handle many requests. This wrapper owns: - SparkTTSModel (vLLM engine) - IndicPromptBuilder - BiCodecDecoder - SparkTTSPipeline + Veena3SlidingWindowPipeline - optional SuperResolutionService Imports framework-agnostic inference + processing code from `veena3modal/core`, `veena3modal/processing`, and `veena3modal/audio`. """ from __future__ import annotations import os import sys import time import logging from dataclasses import dataclass, field from typing import Any, Dict, List, Optional, Tuple logger = logging.getLogger(__name__) @dataclass class TTSRuntime: """ Holds long-lived, per-container inference objects. Thread Safety: This class is designed for async concurrency within a single container. The vLLM engine handles internal batching and scheduling. Do NOT share across processes without proper synchronization. """ model: Any = None pipeline: Any = None streaming_pipeline: Any = None long_text_processor: Any = None prompt_builder: Any = None bicodec_decoder: Any = None sr_service: Optional[Any] = None model_version: str = "not_loaded" is_loaded: bool = False load_time_ms: float = 0.0 # Configuration model_path: str = "" bicodec_path: str = "" sr_checkpoint_dir: Optional[str] = None device: str = "cuda" # OPTIMIZATION: Pre-computed global tokens per speaker (eliminates ~110ms pre-roll in streaming) # Maps speaker_name -> list of 32 global token IDs captured at startup speaker_global_cache: Dict[str, List[int]] = field(default_factory=dict) # Module-level singleton (per-container) _runtime: Optional[TTSRuntime] = None def get_runtime() -> Optional[TTSRuntime]: """Get the current TTS runtime singleton.""" return _runtime def is_initialized() -> bool: """Check if the runtime is initialized and ready.""" return _runtime is not None and _runtime.is_loaded def initialize_runtime( model_path: Optional[str] = None, bicodec_path: Optional[str] = None, sr_checkpoint_dir: Optional[str] = None, device: str = "cuda", hf_token: Optional[str] = None, gpu_memory_utilization: float = 0.25, enable_sr: bool = False, num_engines: int = 1, max_num_batched_tokens: Optional[int] = None, max_num_seqs: Optional[int] = None, enable_chunked_prefill: Optional[bool] = None, enable_prefix_caching: Optional[bool] = None, disable_log_stats: Optional[bool] = None, enforce_eager: Optional[bool] = None, precompute_speaker_globals: Optional[bool] = None, ) -> TTSRuntime: """ Initialize the TTS runtime with all components. This should be called once per container (e.g., in Modal's @modal.enter). Args: model_path: Path to Spark TTS model (env: SPARK_TTS_MODEL_PATH) bicodec_path: Path to BiCodec model (env: BICODEC_MODEL_PATH, defaults to model_path) sr_checkpoint_dir: Path to super-resolution checkpoints (env: AP_BWE_CHECKPOINT_DIR) device: Device for inference (cuda/cpu) hf_token: HuggingFace token for private models (env: HF_TOKEN) gpu_memory_utilization: vLLM GPU memory fraction (default: 0.25) enable_sr: Enable super-resolution service num_engines: Number of vLLM engine instances (default: 1, set to 2-3 for Tier 3 optimization) max_num_batched_tokens: Optional vLLM scheduler cap override. max_num_seqs: Optional vLLM concurrent sequence cap override. enable_chunked_prefill: Optional vLLM chunked prefill toggle. enable_prefix_caching: Optional vLLM prefix caching toggle. disable_log_stats: Optional vLLM internal stats log toggle. enforce_eager: Optional vLLM eager-mode toggle (disables CUDA graphs when true). precompute_speaker_globals: Whether to warm speaker global tokens at startup. If None, reads PRECOMPUTE_SPEAKER_GLOBALS env var (default: false). Returns: Initialized TTSRuntime instance Raises: RuntimeError: If initialization fails """ global _runtime start_time = time.time() # Resolve paths from env vars if not provided model_path = model_path or os.environ.get( 'SPARK_TTS_MODEL_PATH', os.environ.get('MODEL_PATH', '/models/spark_tts_4speaker') ) bicodec_path = bicodec_path or os.environ.get( 'BICODEC_MODEL_PATH', model_path # BiCodec is usually in the same directory ) sr_checkpoint_dir = sr_checkpoint_dir or os.environ.get('AP_BWE_CHECKPOINT_DIR') hf_token = hf_token or os.environ.get('HF_TOKEN') or os.environ.get('HUGGING_FACE_HUB_TOKEN') if precompute_speaker_globals is None: raw_flag = os.environ.get("PRECOMPUTE_SPEAKER_GLOBALS", "false").strip().lower() precompute_speaker_globals = raw_flag in {"1", "true", "yes", "on"} logger.info(f"Initializing TTS runtime...") logger.info(f" Model path: {model_path}") logger.info(f" BiCodec path: {bicodec_path}") logger.info(f" SR enabled: {enable_sr}, path: {sr_checkpoint_dir}") try: # Import inference components (vendored into veena3modal for Modal deployment) from veena3modal.core.model_loader import SparkTTSModel from veena3modal.core.pipeline import SparkTTSPipeline from veena3modal.core.bicodec_decoder import BiCodecDecoder from veena3modal.processing.prompt_builder import IndicPromptBuilder from veena3modal.processing.long_text_processor import LongTextProcessor engine_kwargs: Dict[str, Any] = {} if isinstance(max_num_batched_tokens, int) and max_num_batched_tokens > 0: engine_kwargs["max_num_batched_tokens"] = max_num_batched_tokens if isinstance(max_num_seqs, int) and max_num_seqs > 0: engine_kwargs["max_num_seqs"] = max_num_seqs if enable_chunked_prefill is not None: engine_kwargs["enable_chunked_prefill"] = bool(enable_chunked_prefill) if enable_prefix_caching is not None: engine_kwargs["enable_prefix_caching"] = bool(enable_prefix_caching) if disable_log_stats is not None: engine_kwargs["disable_log_stats"] = bool(disable_log_stats) if enforce_eager is not None: engine_kwargs["enforce_eager"] = bool(enforce_eager) # Load SparkTTS model with vLLM engine if num_engines > 1: # TIER 3 OPTIMIZATION: Multiple vLLM engines on same GPU # Each engine gets gpu_memory_utilization / num_engines to share GPU fairly. # E.g., 3 engines at 0.08 each = 0.24 total, ~8GB per engine, ~24GB total. from veena3modal.core.multi_engine import create_multi_engine_model per_engine_mem = gpu_memory_utilization / num_engines logger.info(f"Loading {num_engines} vLLM engines ({per_engine_mem:.2f} GPU mem each)...") model = create_multi_engine_model( model_path=model_path, num_engines=num_engines, hf_token=hf_token, gpu_memory_per_engine=per_engine_mem, **engine_kwargs, ) else: logger.info("Loading SparkTTS model with vLLM...") model = SparkTTSModel( model_path=model_path, hf_token=hf_token, gpu_memory_utilization=gpu_memory_utilization, **engine_kwargs, ) # Initialize prompt builder logger.info("Initializing prompt builder...") prompt_builder = IndicPromptBuilder( tokenizer=model.tokenizer, model=model, ) # Initialize BiCodec decoder logger.info("Initializing BiCodec decoder...") bicodec_decoder = BiCodecDecoder( device=device, model_path=bicodec_path, ) # Initialize pipeline logger.info("Initializing TTS pipeline...") pipeline = SparkTTSPipeline( model=model, prompt_builder=prompt_builder, bicodec_decoder=bicodec_decoder, ) # Initialize streaming pipeline (for M4) streaming_pipeline = None try: from veena3modal.core.streaming_pipeline import Veena3SlidingWindowPipeline logger.info("Initializing streaming pipeline...") # NOTE: Parameter is named 'snac_decoder' for legacy reasons, but works with BiCodecDecoder streaming_pipeline = Veena3SlidingWindowPipeline( model=model, prompt_builder=prompt_builder, snac_decoder=bicodec_decoder, # BiCodecDecoder is interface-compatible ) except ImportError as e: logger.warning(f"Streaming pipeline not available: {e}") # Initialize long text processor once (avoid per-request construction/logging) long_text_processor = LongTextProcessor( pipeline=pipeline, streaming_pipeline=streaming_pipeline, ) # Initialize super-resolution (optional) sr_service = None if enable_sr and sr_checkpoint_dir: try: from veena3modal.core.super_resolution import SuperResolutionService logger.info(f"Initializing super-resolution from {sr_checkpoint_dir}...") sr_service = SuperResolutionService(checkpoint_dir=sr_checkpoint_dir) # Load the model explicitly if sr_service.load_model(device=device): logger.info("✅ Super-resolution model loaded successfully") else: logger.warning("Super-resolution model failed to load") sr_service = None except Exception as e: logger.warning(f"Super-resolution not available: {e}") # Determine model version model_version = os.path.basename(model_path.rstrip('/')) if not model_version: model_version = "spark-tts" load_time = (time.time() - start_time) * 1000 # Create runtime _runtime = TTSRuntime( model=model, pipeline=pipeline, streaming_pipeline=streaming_pipeline, long_text_processor=long_text_processor, prompt_builder=prompt_builder, bicodec_decoder=bicodec_decoder, sr_service=sr_service, model_version=model_version, is_loaded=True, load_time_ms=load_time, model_path=model_path, bicodec_path=bicodec_path, sr_checkpoint_dir=sr_checkpoint_dir, device=device, ) logger.info(f"✅ TTS runtime initialized in {load_time:.0f}ms") logger.info(f" Model version: {model_version}") # Pre-compute speaker globals only when explicitly enabled. # Startup warmup runs async generation and can bind AsyncLLMEngine state # to a non-serving event loop if executed in a different loop context. if precompute_speaker_globals and num_engines <= 1: logger.info("Speaker globals startup precompute enabled") _precompute_speaker_globals(_runtime) elif num_engines > 1: logger.info("Speaker globals cache startup precompute skipped (multi-engine mode)") else: logger.info("Speaker globals cache startup precompute skipped (disabled)") # Update FastAPI app with model version try: from veena3modal.api.fastapi_app import set_model_version set_model_version(model_version) except ImportError: pass return _runtime except Exception as e: logger.error(f"❌ Failed to initialize TTS runtime: {e}") import traceback traceback.print_exc() raise RuntimeError(f"TTS runtime initialization failed: {e}") from e def _precompute_speaker_globals(runtime: TTSRuntime) -> None: """ Pre-compute and cache global tokens for all 12 speakers at startup. OPTIMIZATION: BiCodec streaming requires 32 "global tokens" before any audio can be emitted. These encode speaker identity via FSQ quantization. Since we have only 12 fixed speakers, we generate one short utterance per speaker at startup, capture the 32 global tokens, and cache them. During streaming, we inject cached globals via build_prefix_with_globals(), skipping the ~110ms global token pre-roll phase entirely. This runs synchronously at startup (adds ~5-10s to cold start, saves ~110ms per streaming TTFB). """ import asyncio import re from vllm import SamplingParams from veena3modal.core.constants import ( SPEAKER_MAP, TRAINING_STOP_TOKEN_IDS, DEFAULT_TEMPERATURE, DEFAULT_TOP_K, DEFAULT_TOP_P, ) logger.info("Pre-computing global tokens for all speakers...") t_start = time.time() test_text = "Hello, this is a voice test." sampling_params = SamplingParams( temperature=DEFAULT_TEMPERATURE, top_k=DEFAULT_TOP_K, top_p=DEFAULT_TOP_P, max_tokens=128, # Only need ~60 tokens (32 global + some semantic) stop=TRAINING_STOP_TOKEN_IDS, skip_special_tokens=False, ) async def _capture_globals_for_speaker(speaker_name: str) -> Optional[List[int]]: """Generate a short utterance and capture the 32 global tokens.""" prompt = runtime.prompt_builder.build_prefix(speaker_name, test_text) request_id = f"warmup-{speaker_name}-{int(time.time() * 1000)}" final_output = None async for request_output in runtime.model.engine.generate( prompt=prompt, sampling_params=sampling_params, request_id=request_id, ): final_output = request_output if final_output is None: return None generated_text = final_output.outputs[0].text global_matches = re.findall(r"<\|bicodec_global_(\d+)\|>", generated_text) global_ids = [int(t) for t in global_matches] if len(global_ids) >= 32: return global_ids[:32] return None async def _run_all(): for speaker_name in SPEAKER_MAP.keys(): try: global_ids = await _capture_globals_for_speaker(speaker_name) if global_ids and len(global_ids) == 32: runtime.speaker_global_cache[speaker_name] = global_ids logger.info(f" Cached globals for {speaker_name}") else: logger.warning(f" Failed to capture globals for {speaker_name}") except Exception as e: logger.warning(f" Error caching globals for {speaker_name}: {e}") # Run async warmup in event loop try: loop = asyncio.get_event_loop() if loop.is_running(): # Already in async context (unlikely at startup, but handle it) import concurrent.futures with concurrent.futures.ThreadPoolExecutor() as pool: pool.submit(lambda: asyncio.run(_run_all())).result() else: loop.run_until_complete(_run_all()) except RuntimeError: # No event loop exists yet asyncio.run(_run_all()) elapsed = (time.time() - t_start) * 1000 cached = len(runtime.speaker_global_cache) total = len(SPEAKER_MAP) logger.info(f"Speaker globals cached: {cached}/{total} speakers in {elapsed:.0f}ms") def _get_long_text_processor(runtime: TTSRuntime): """Return cached LongTextProcessor, creating it lazily if needed.""" if runtime.long_text_processor is None: from veena3modal.processing.long_text_processor import LongTextProcessor runtime.long_text_processor = LongTextProcessor( pipeline=runtime.pipeline, streaming_pipeline=runtime.streaming_pipeline, ) return runtime.long_text_processor async def generate_speech( text: str, speaker: str, temperature: float = 0.8, top_k: int = 50, top_p: float = 1.0, max_tokens: int = 4096, repetition_penalty: float = 1.05, seed: Optional[int] = None, output_sample_rate: str = "16khz", ) -> Tuple[Optional[bytes], Dict[str, Any]]: """ Generate speech audio (non-streaming). Args: text: Text to synthesize (already normalized) speaker: Internal speaker name (resolved) temperature: Sampling temperature top_k: Top-k sampling top_p: Nucleus sampling max_tokens: Maximum tokens to generate repetition_penalty: Repetition penalty seed: Random seed for reproducibility output_sample_rate: "16khz" or "48khz" (triggers super-resolution) Returns: Tuple of (audio_bytes, metrics_dict) audio_bytes: WAV audio bytes (16kHz or 48kHz, 16-bit PCM) or None if failed metrics_dict: Dictionary with timing metrics Raises: RuntimeError: If runtime not initialized """ if not is_initialized(): raise RuntimeError("TTS runtime not initialized") runtime = get_runtime() metrics = { "ttfb_ms": 0, "generation_ms": 0, "tokens_generated": 0, "audio_duration_seconds": 0.0, "sr_applied": False, "output_sample_rate": 16000, } start_time = time.time() try: # Generate audio using pipeline (always at 16kHz) audio_bytes, perf = await runtime.pipeline.generate_speech_profiled( speaker=speaker, text=text, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, ) if perf: metrics.update(perf) generation_time = time.time() - start_time metrics["generation_ms"] = int(generation_time * 1000) metrics["ttfb_ms"] = metrics["generation_ms"] # Non-streaming: TTFB = total time if audio_bytes: # Apply super-resolution if requested and available sample_rate = 16000 if output_sample_rate == "48khz" or logger.isEnabledFor(logging.DEBUG): logger.debug( "SR check: output_sample_rate=%r, sr_service=%s, is_loaded=%s", output_sample_rate, runtime.sr_service is not None, runtime.sr_service.is_loaded if runtime.sr_service else False, ) if output_sample_rate == "48khz" and runtime.sr_service and runtime.sr_service.is_loaded: try: sr_start = time.time() logger.info(f"Applying super-resolution to {len(audio_bytes)} bytes...") audio_bytes = _apply_super_resolution(audio_bytes, runtime.sr_service) sr_time = (time.time() - sr_start) * 1000 metrics["sr_ms"] = int(sr_time) metrics["sr_applied"] = True sample_rate = 48000 logger.info(f"✅ Super-resolution applied in {sr_time:.1f}ms, output={len(audio_bytes)} bytes") except Exception as e: logger.warning(f"Super-resolution failed, returning 16kHz: {e}") import traceback traceback.print_exc() metrics["output_sample_rate"] = sample_rate # Calculate audio duration audio_duration = (len(audio_bytes) - 44) / (sample_rate * 2) # -44 for WAV header metrics["audio_duration_seconds"] = max(0.0, audio_duration) return audio_bytes, metrics except Exception as e: logger.error(f"Speech generation failed: {e}") raise def _apply_super_resolution(audio_bytes: bytes, sr_service) -> bytes: """ Apply super-resolution to audio bytes. Args: audio_bytes: WAV audio at 16kHz sr_service: SuperResolutionService instance Returns: WAV audio at 48kHz """ import numpy as np import struct import torch # Parse WAV header and extract PCM data if len(audio_bytes) < 44: raise ValueError("Invalid WAV data") # Extract PCM data (skip 44-byte WAV header) pcm_data = np.frombuffer(audio_bytes[44:], dtype=np.int16) # Convert to float32 for SR model audio_float = pcm_data.astype(np.float32) / 32768.0 # Convert numpy to torch tensor for process_chunk audio_tensor = torch.from_numpy(audio_float) # Apply super-resolution using process_chunk (16kHz -> 48kHz) # process_chunk expects [batch, samples] or [samples] and returns [batch, samples] upsampled_tensor = sr_service.process_chunk(audio_tensor) # Convert back to numpy, squeeze batch dim if present upsampled = upsampled_tensor.squeeze().cpu().numpy() # Convert back to int16 upsampled_int16 = np.clip(upsampled * 32768.0, -32768, 32767).astype(np.int16) # Create new WAV header for 48kHz sample_rate = 48000 num_channels = 1 bits_per_sample = 16 byte_rate = sample_rate * num_channels * bits_per_sample // 8 block_align = num_channels * bits_per_sample // 8 data_size = len(upsampled_int16) * 2 wav_header = struct.pack( '<4sI4s4sIHHIIHH4sI', b'RIFF', 36 + data_size, b'WAVE', b'fmt ', 16, # fmt chunk size 1, # PCM format num_channels, sample_rate, byte_rate, block_align, bits_per_sample, b'data', data_size, ) return wav_header + upsampled_int16.tobytes() async def generate_speech_chunked( text: str, speaker: str, temperature: float = 0.8, top_k: int = 50, top_p: float = 1.0, max_tokens: int = 4096, repetition_penalty: float = 1.05, seed: Optional[int] = None, sample_rate: int = 16000, output_sample_rate: str = "16khz", ) -> Tuple[Optional[bytes], Dict[str, Any]]: """ Generate speech with automatic text chunking for long inputs. Uses LongTextProcessor to split text and stitch audio. Args: text: Text to synthesize speaker: Internal speaker name ... (same as generate_speech) sample_rate: Internal generation sample rate (always 16000) output_sample_rate: "16khz" or "48khz" (triggers super-resolution) Returns: Tuple of (audio_bytes, metrics_dict) """ if not is_initialized(): raise RuntimeError("TTS runtime not initialized") runtime = get_runtime() metrics = { "ttfb_ms": 0, "generation_ms": 0, "chunks_processed": 0, "audio_duration_seconds": 0.0, "text_chunked": False, "sr_applied": False, "output_sample_rate": 16000, } start_time = time.time() try: long_processor = _get_long_text_processor(runtime) # Check if chunking is needed if long_processor.should_chunk(text): metrics["text_chunked"] = True audio_bytes = await long_processor.generate_with_chunking( text=text, speaker=speaker, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, sample_rate=sample_rate, ) else: # Short text: direct generation audio_bytes, perf = await runtime.pipeline.generate_speech_profiled( speaker=speaker, text=text, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, ) if perf: metrics.update(perf) generation_time = time.time() - start_time metrics["generation_ms"] = int(generation_time * 1000) metrics["ttfb_ms"] = metrics["generation_ms"] if audio_bytes: # Apply super-resolution if requested and available final_sample_rate = 16000 if output_sample_rate == "48khz" and runtime.sr_service and runtime.sr_service.is_loaded: try: sr_start = time.time() logger.info(f"Applying super-resolution to chunked audio ({len(audio_bytes)} bytes)...") audio_bytes = _apply_super_resolution(audio_bytes, runtime.sr_service) sr_time = (time.time() - sr_start) * 1000 metrics["sr_ms"] = int(sr_time) metrics["sr_applied"] = True final_sample_rate = 48000 logger.info(f"✅ Super-resolution applied in {sr_time:.1f}ms, output={len(audio_bytes)} bytes") except Exception as e: logger.warning(f"Super-resolution failed, returning 16kHz: {e}") import traceback traceback.print_exc() metrics["output_sample_rate"] = final_sample_rate audio_duration = (len(audio_bytes) - 44) / (final_sample_rate * 2) metrics["audio_duration_seconds"] = max(0.0, audio_duration) return audio_bytes, metrics except Exception as e: logger.error(f"Chunked speech generation failed: {e}") raise from typing import AsyncGenerator async def generate_speech_streaming( text: str, speaker: str, temperature: float = 0.8, top_k: int = 50, top_p: float = 1.0, max_tokens: int = 4096, repetition_penalty: float = 1.05, seed: Optional[int] = None, enable_chunking: bool = True, ) -> AsyncGenerator[Tuple[bytes, Dict[str, Any]], None]: """ Generate speech audio with true streaming (yields chunks as they're generated). This is the core streaming implementation for M4. First yield includes WAV header + first PCM chunk. Subsequent yields are raw PCM chunks. Args: text: Text to synthesize (already normalized) speaker: Internal speaker name (resolved) temperature: Sampling temperature top_k: Top-k sampling top_p: Nucleus sampling max_tokens: Maximum tokens to generate repetition_penalty: Repetition penalty seed: Random seed for reproducibility enable_chunking: Enable text chunking for long inputs with voice consistency Yields: Tuple of (audio_bytes, metrics_dict) - First yield: WAV header (44 bytes) prepended to first PCM chunk - Subsequent yields: Raw PCM chunks (int16, 16kHz) - Final yield: metrics_dict has final timing info Raises: RuntimeError: If runtime not initialized or streaming pipeline unavailable """ if not is_initialized(): raise RuntimeError("TTS runtime not initialized") runtime = get_runtime() if runtime.streaming_pipeline is None: raise RuntimeError("Streaming pipeline not available") # Import audio utils for WAV header from veena3modal.audio.utils import create_wav_header # Metrics tracking metrics = { "ttfb_ms": 0, "chunks_sent": 0, "total_bytes": 0, "audio_duration_seconds": 0.0, "text_chunked": False, } start_time = time.time() first_chunk_time = None total_pcm_bytes = 0 sample_rate = 16000 # BiCodec sample rate # Check if we need text chunking long_processor = _get_long_text_processor(runtime) needs_chunking = enable_chunking and long_processor.should_chunk(text) if needs_chunking: # Chunked streaming with voice consistency (global token caching) metrics["text_chunked"] = True async for audio_chunk, chunk_metrics in _stream_chunked_text( runtime, long_processor, text, speaker, temperature, top_k, top_p, max_tokens, repetition_penalty, seed, sample_rate ): # First chunk: prepend WAV header if first_chunk_time is None: first_chunk_time = time.time() metrics["ttfb_ms"] = int((first_chunk_time - start_time) * 1000) # Create streaming WAV header (size=0 for unknown length) wav_header = create_wav_header(sample_rate=sample_rate, data_size=0) audio_chunk = wav_header + audio_chunk total_pcm_bytes += len(audio_chunk) - (44 if metrics["chunks_sent"] == 0 else 0) metrics["chunks_sent"] += 1 metrics["total_bytes"] = total_pcm_bytes + 44 # Include header metrics["audio_duration_seconds"] = total_pcm_bytes / (sample_rate * 2) yield audio_chunk, metrics else: # Simple streaming (no text chunking) # OPTIMIZATION: If we have pre-cached global tokens for this speaker, # use continuation mode to skip the ~110ms global token pre-roll. # The model jumps straight to semantic token generation. cached_globals = runtime.speaker_global_cache.get(speaker) if cached_globals: # Fast path: use cached globals, skip global token generation entirely stream_gen = runtime.streaming_pipeline.generate_speech_stream_indic_continuation( speaker=speaker, text=text, global_ids=cached_globals, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, emit_progress=True, ) else: # Fallback: no cached globals, generate them inline (original path) stream_gen = runtime.streaming_pipeline.generate_speech_stream_indic( speaker=speaker, text=text, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, emit_progress=True, ) async for stream_item in stream_gen: stream_metrics: Dict[str, Any] = {} audio_chunk = stream_item if ( isinstance(stream_item, tuple) and len(stream_item) == 2 and isinstance(stream_item[1], dict) ): audio_chunk = stream_item[0] stream_metrics = stream_item[1] if stream_metrics: metrics.update(stream_metrics) if not audio_chunk: # Metrics-only event from streaming pipeline. continue # First chunk: prepend WAV header if first_chunk_time is None: first_chunk_time = time.time() metrics["ttfb_ms"] = int((first_chunk_time - start_time) * 1000) wav_header = create_wav_header(sample_rate=sample_rate, data_size=0) audio_chunk = wav_header + audio_chunk total_pcm_bytes += len(audio_chunk) - (44 if metrics["chunks_sent"] == 0 else 0) metrics["chunks_sent"] += 1 metrics["total_bytes"] = total_pcm_bytes + 44 metrics["audio_duration_seconds"] = total_pcm_bytes / (sample_rate * 2) yield audio_chunk, metrics async def _stream_chunked_text( runtime: TTSRuntime, long_processor, text: str, speaker: str, temperature: float, top_k: int, top_p: float, max_tokens: int, repetition_penalty: float, seed: Optional[int], sample_rate: int, ) -> AsyncGenerator[Tuple[bytes, Dict[str, Any]], None]: """ Internal helper: stream chunked text with voice consistency. Uses global token caching from first chunk to maintain voice across chunks. """ from veena3modal.audio.crossfade import crossfade_bytes_int16 # Chunk text using the long text processor's chunker chunks = long_processor.chunk_text(text) if not chunks: return captured_globals: Optional[List[int]] = None previous_chunk_tail: Optional[bytes] = None chunk_metrics = {"chunks_processed": 0} for i, chunk_text in enumerate(chunks): chunk_metrics["chunks_processed"] = i + 1 if i == 0: # First chunk: capture global tokens for voice consistency async for audio_bytes, global_ids in runtime.streaming_pipeline.generate_speech_stream_indic_first_chunk( speaker=speaker, text=chunk_text, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, ): if captured_globals is None and global_ids: captured_globals = global_ids # Crossfade with previous tail (inter-chunk stitching) to_emit, previous_chunk_tail = crossfade_bytes_int16( previous_chunk_tail, audio_bytes, sample_rate_hz=sample_rate, crossfade_ms=50, ) if to_emit: yield to_emit, chunk_metrics else: # Continuation chunks: use captured globals for voice consistency if captured_globals is None: logger.warning(f"No captured globals for chunk {i+1}, using regular streaming") async for audio_bytes in runtime.streaming_pipeline.generate_speech_stream_indic( speaker=speaker, text=chunk_text, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, ): to_emit, previous_chunk_tail = crossfade_bytes_int16( previous_chunk_tail, audio_bytes, sample_rate_hz=sample_rate, crossfade_ms=50, ) if to_emit: yield to_emit, chunk_metrics else: # Use continuation method with cached globals async for audio_bytes in runtime.streaming_pipeline.generate_speech_stream_indic_continuation( speaker=speaker, text=chunk_text, global_ids=captured_globals, temperature=temperature, top_k=top_k, top_p=top_p, max_tokens=max_tokens, repetition_penalty=repetition_penalty, seed=seed, ): to_emit, previous_chunk_tail = crossfade_bytes_int16( previous_chunk_tail, audio_bytes, sample_rate_hz=sample_rate, crossfade_ms=50, ) if to_emit: yield to_emit, chunk_metrics # Flush remaining tail if previous_chunk_tail: yield previous_chunk_tail, chunk_metrics