lattifai 1.0.4__py3-none-any.whl → 1.0.5__py3-none-any.whl

lattifai/cli/app_installer.py

@@ -5,6 +5,8 @@ import subprocess
 import sys
 from pathlib import Path
 
+from lattifai.utils import safe_print
+
 
 def check_command_exists(cmd: str) -> bool:
     """Check if a command exists in PATH."""
@@ -19,17 +21,17 @@ def install_nodejs():
     """Install Node.js based on the operating system."""
     system = platform.system().lower()
 
-    print("📦 Node.js not found. Installing Node.js...\n")
+    safe_print("📦 Node.js not found. Installing Node.js...\n")
 
     try:
         if system == "darwin":  # macOS
             # Check if Homebrew is installed
             if check_command_exists("brew"):
-                print("🍺 Using Homebrew to install Node.js...")
+                safe_print("🍺 Using Homebrew to install Node.js...")
                 subprocess.run(["brew", "install", "node"], check=True)
-                print("✓ Node.js installed via Homebrew\n")
+                safe_print("✓ Node.js installed via Homebrew\n")
             else:
-                print("❌ Homebrew not found.")
+                safe_print("❌ Homebrew not found.")
                 print("   Please install Homebrew first:")
                 print(
                     '   /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"'
@@ -40,46 +42,46 @@ def install_nodejs():
         elif system == "linux":
             # Try common package managers
             if check_command_exists("apt"):
-                print("🐧 Using apt to install Node.js...")
+                safe_print("🐧 Using apt to install Node.js...")
                 subprocess.run(["sudo", "apt", "update"], check=True)
                 subprocess.run(["sudo", "apt", "install", "-y", "nodejs", "npm"], check=True)
-                print("✓ Node.js installed via apt\n")
+                safe_print("✓ Node.js installed via apt\n")
             elif check_command_exists("yum"):
-                print("🐧 Using yum to install Node.js...")
+                safe_print("🐧 Using yum to install Node.js...")
                 subprocess.run(["sudo", "yum", "install", "-y", "nodejs", "npm"], check=True)
-                print("✓ Node.js installed via yum\n")
+                safe_print("✓ Node.js installed via yum\n")
             elif check_command_exists("dnf"):
-                print("🐧 Using dnf to install Node.js...")
+                safe_print("🐧 Using dnf to install Node.js...")
                 subprocess.run(["sudo", "dnf", "install", "-y", "nodejs", "npm"], check=True)
-                print("✓ Node.js installed via dnf\n")
+                safe_print("✓ Node.js installed via dnf\n")
             elif check_command_exists("pacman"):
-                print("🐧 Using pacman to install Node.js...")
+                safe_print("🐧 Using pacman to install Node.js...")
                 subprocess.run(["sudo", "pacman", "-S", "--noconfirm", "nodejs", "npm"], check=True)
-                print("✓ Node.js installed via pacman\n")
+                safe_print("✓ Node.js installed via pacman\n")
             else:
-                print("❌ No supported package manager found (apt/yum/dnf/pacman).")
+                safe_print("❌ No supported package manager found (apt/yum/dnf/pacman).")
                 print("   Please install Node.js manually from: https://nodejs.org/")
                 sys.exit(1)
 
         elif system == "windows":
-            print("❌ Automatic installation on Windows is not supported.")
+            safe_print("❌ Automatic installation on Windows is not supported.")
             print("   Please download and install Node.js from: https://nodejs.org/")
             print("   Then run this command again.")
             sys.exit(1)
 
         else:
-            print(f"❌ Unsupported operating system: {system}")
+            safe_print(f"❌ Unsupported operating system: {system}")
             print("   Please install Node.js manually from: https://nodejs.org/")
             sys.exit(1)
 
         # Verify installation
         if not check_command_exists("npm"):
-            print("❌ Node.js installation verification failed.")
+            safe_print("❌ Node.js installation verification failed.")
             print("   Please restart your terminal and try again.")
             sys.exit(1)
 
     except subprocess.CalledProcessError as e:
-        print(f"\n❌ Error during Node.js installation: {e}")
+        safe_print(f"\n❌ Error during Node.js installation: {e}")
         print("   Please install Node.js manually from: https://nodejs.org/")
         sys.exit(1)
 
@@ -90,49 +92,49 @@ def main():
     app_dir = Path(__file__).parent.parent.parent.parent / "app"
 
     if not app_dir.exists():
-        print(f"❌ Error: app directory not found at {app_dir}")
+        safe_print(f"❌ Error: app directory not found at {app_dir}")
         print("   Make sure you're in the lattifai-python repository.")
         sys.exit(1)
 
-    print("🚀 Installing lai-app (LattifAI Web Application)...\n")
+    safe_print("🚀 Installing lai-app (LattifAI Web Application)...\n")
 
     # Check if npm is installed, if not, install Node.js
     if not check_command_exists("npm"):
         install_nodejs()
     else:
         npm_version = subprocess.run(["npm", "--version"], capture_output=True, text=True, check=True).stdout.strip()
-        print(f"✓ npm is already installed (v{npm_version})\n")
+        safe_print(f"✓ npm is already installed (v{npm_version})\n")
 
     # Change to app directory and run installation
     try:
-        print(f"📁 Working directory: {app_dir}\n")
+        safe_print(f"📁 Working directory: {app_dir}\n")
 
         # Install dependencies
-        print("📦 Installing dependencies...")
+        safe_print("📦 Installing dependencies...")
         subprocess.run(["npm", "install"], cwd=app_dir, check=True)
-        print("✓ Dependencies installed\n")
+        safe_print("✓ Dependencies installed\n")
 
         # Build the application
-        print("🔨 Building application...")
+        safe_print("🔨 Building application...")
         subprocess.run(["npm", "run", "build"], cwd=app_dir, check=True)
-        print("✓ Application built\n")
+        safe_print("✓ Application built\n")
 
         # Link globally
-        print("🔗 Linking lai-app command globally...")
+        safe_print("🔗 Linking lai-app command globally...")
         subprocess.run(["npm", "link"], cwd=app_dir, check=True)
-        print("✓ lai-app command linked globally\n")
+        safe_print("✓ lai-app command linked globally\n")
 
-        print("=" * 60)
-        print("✅ lai-app installed successfully!")
-        print("=" * 60)
-        print("\n🎉 You can now run:")
+        safe_print("=" * 60)
+        safe_print("✅ lai-app installed successfully!")
+        safe_print("=" * 60)
+        safe_print("\n🎉 You can now run:")
         print("   lai-app              # Start the web application")
         print("   lai-app --help       # Show help")
         print("   lai-app --port 8080  # Use custom port")
-        print("\n📖 For more information, see app/CLI_USAGE.md\n")
+        safe_print("\n📖 For more information, see app/CLI_USAGE.md\n")
 
     except subprocess.CalledProcessError as e:
-        print(f"\n❌ Error during installation: {e}")
+        safe_print(f"\n❌ Error during installation: {e}")
         sys.exit(1)
 
 

lattifai/cli/caption.py

@@ -7,6 +7,7 @@ from lhotse.utils import Pathlike
 from typing_extensions import Annotated
 
 from lattifai.config import CaptionConfig
+from lattifai.utils import safe_print
 
 
 @run.cli.entrypoint(name="convert", namespace="caption")
@@ -55,7 +56,7 @@ def convert(
     caption = Caption.read(input_path, normalize_text=normalize_text)
     caption.write(output_path, include_speaker_in_text=include_speaker_in_text)
 
-    print(f"✅ Converted {input_path} -> {output_path}")
+    safe_print(f"✅ Converted {input_path} -> {output_path}")
     return output_path
 
 
@@ -63,7 +64,6 @@ def convert(
 def normalize(
     input_path: Pathlike,
     output_path: Pathlike,
-    caption: Annotated[Optional[CaptionConfig], run.Config[CaptionConfig]] = None,
 ):
     """
     Normalize caption text by cleaning HTML entities and whitespace.
@@ -81,9 +81,6 @@ def normalize(
     Args:
         input_path: Path to input caption file to normalize
         output_path: Path to output caption file (defaults to overwriting input file)
-        caption: Caption configuration for text normalization.
-            Fields: input_format, output_format, normalize_text (automatically enabled),
-                    encoding
 
     Examples:
         # Normalize and save to new file (positional arguments)
@@ -92,13 +89,9 @@ def normalize(
         # Normalize with format conversion
         lai caption normalize input.vtt output.srt
 
-        # Normalize with custom caption config
-        lai caption normalize input.srt output.srt \\
-            caption.encoding=utf-8
-
         # Using keyword arguments (traditional syntax)
-        lai caption normalize \\
-            input_path=input.srt \\
+        lai caption normalize \
+            input_path=input.srt \
             output_path=output.srt
     """
     from pathlib import Path
@@ -112,9 +105,9 @@ def normalize(
     caption_obj.write(output_path, include_speaker_in_text=True)
 
     if output_path == input_path:
-        print(f"✅ Normalized {input_path} (in-place)")
+        safe_print(f"✅ Normalized {input_path} (in-place)")
     else:
-        print(f"✅ Normalized {input_path} -> {output_path}")
+        safe_print(f"✅ Normalized {input_path} -> {output_path}")
 
     return output_path
 
@@ -124,7 +117,6 @@ def shift(
     input_path: Pathlike,
     output_path: Pathlike,
     seconds: float,
-    caption: Annotated[Optional[CaptionConfig], run.Config[CaptionConfig]] = None,
 ):
     """
     Shift caption timestamps by a specified number of seconds.
@@ -140,8 +132,6 @@ def shift(
         output_path: Path to output caption file (can be same as input for in-place modification)
         seconds: Number of seconds to shift timestamps. Positive values delay captions,
                  negative values advance them earlier.
-        caption: Caption configuration for reading/writing.
-            Fields: input_format, output_format, encoding
 
     Examples:
         # Delay captions by 2 seconds (positional arguments)
@@ -181,9 +171,9 @@ def shift(
         direction = f"advanced by {abs(seconds)}s"
 
     if output_path == input_path:
-        print(f"✅ Shifted timestamps {direction} in {input_path} (in-place)")
+        safe_print(f"✅ Shifted timestamps {direction} in {input_path} (in-place)")
     else:
-        print(f"✅ Shifted timestamps {direction}: {input_path} -> {output_path}")
+        safe_print(f"✅ Shifted timestamps {direction}: {input_path} -> {output_path}")
 
     return output_path
 

lattifai/cli/server.py

@@ -4,6 +4,8 @@ import os
 import colorful
 import uvicorn
 
+from lattifai.utils import safe_print
+
 
 def main():
     """Launch the LattifAI Web Interface."""
@@ -29,7 +31,7 @@ def main():
 
     args = parser.parse_args()
 
-    print(colorful.bold_green("🚀 Launching LattifAI Backend Server..."))
+    safe_print(colorful.bold_green("🚀 Launching LattifAI Backend Server..."))
     print(colorful.cyan(f"Server running at http://localhost:{args.port}"))
     print(colorful.yellow(f"Host: {args.host}"))
     print(colorful.yellow(f"Auto-reload: {'disabled' if args.no_reload else 'enabled'}"))

lattifai/cli/transcribe.py

@@ -3,10 +3,8 @@
 from typing import Optional
 
 import nemo_run as run
-from lhotse.utils import Pathlike
 from typing_extensions import Annotated
 
-from lattifai.audio2 import AudioLoader, ChannelSelectorType
 from lattifai.cli.alignment import align as alignment_align
 from lattifai.config import (
     AlignmentConfig,
@@ -23,9 +21,8 @@ from lattifai.utils import _resolve_model_path
 def transcribe(
     input: Optional[str] = None,
     output_caption: Optional[str] = None,
-    output_dir: Optional[Pathlike] = None,
-    media_format: str = "mp3",
-    channel_selector: Optional[ChannelSelectorType] = "average",
+    media: Annotated[Optional[MediaConfig], run.Config[MediaConfig]] = None,
+    client: Annotated[Optional[ClientConfig], run.Config[ClientConfig]] = None,
     transcription: Annotated[Optional[TranscriptionConfig], run.Config[TranscriptionConfig]] = None,
 ):
     """
@@ -39,11 +36,8 @@ def transcribe(
     Args:
         input: Path to input audio/video file or YouTube URL (can be provided as positional argument)
         output_caption: Path for output caption file (can be provided as positional argument)
-        output_dir: Directory for output files when using YouTube URL
-        media_format: Media format for YouTube downloads (default: mp3)
-        channel_selector: Audio channel selection strategy (default: average)
-            Options: average, left, right, or an integer channel index.
-            Note: Ignored when input is a URL and Gemini transcriber is used.
+        media: Media configuration for input/output handling.
+            Fields: input_path, output_dir, media_format, channel_selector, streaming_chunk_secs
         transcription: Transcription service configuration.
             Fields: model_name, device, language, gemini_api_key
 
@@ -67,6 +61,11 @@ def transcribe(
         lai transcribe run audio.wav output.srt \\
             transcription.language=zh
 
+        # With MediaConfig settings
+        lai transcribe run audio.wav output.srt \\
+            media.channel_selector=left \\
+            media.streaming_chunk_secs=30.0
+
         # Full configuration with keyword arguments
         lai transcribe run \\
             input=audio.wav \\
@@ -78,68 +77,84 @@ def transcribe(
     from pathlib import Path
 
     import colorful
+    from lattifai_core.client import SyncAPIClient
 
+    from lattifai.audio2 import AudioLoader
     from lattifai.transcription import create_transcriber
+    from lattifai.utils import safe_print
 
-    # Initialize transcription config with defaults
+    # Initialize configs with defaults
+    client_config = client or ClientConfig()
     transcription_config = transcription or TranscriptionConfig()
+    media_config = media or MediaConfig()
+
+    # Initialize client wrapper to properly set client_wrapper
+    client_wrapper = SyncAPIClient(config=client_config)
+    transcription_config.client_wrapper = client_wrapper
+
+    # Initialize client wrapper to properly set client_wrapper
+    client_wrapper = SyncAPIClient(config=client_config)
+    transcription_config.client_wrapper = client_wrapper
 
     # Validate input is required
-    if not input:
-        raise ValueError("Input is required. Provide input as positional argument (file path or URL).")
+    if not input and not media_config.input_path:
+        raise ValueError("Input is required. Provide input as positional argument or media.input_path.")
+
+    # Assign input to media_config if provided
+    if input:
+        media_config.set_input_path(input)
 
     # Detect if input is a URL
-    is_url = input.startswith(("http://", "https://"))
+    is_url = media_config.is_input_remote()
 
     # Prepare output paths
     if is_url:
-        # For URLs, use output_dir
-        if output_dir:
-            output_path = Path(str(output_dir)).expanduser()
-            output_path.mkdir(parents=True, exist_ok=True)
-        else:
-            output_path = Path.cwd()
+        # For URLs, use output_dir from media_config or current directory
+        output_path = media_config.output_dir
     else:
         # For files, use input path directory
-        input_path = Path(str(input))
-        output_path = input_path.parent
+        output_path = Path(media_config.input_path).parent
 
     # Create transcriber
     if not transcription_config.lattice_model_path:
-        transcription_config.lattice_model_path = _resolve_model_path("Lattifai/Lattice-1")
+        transcription_config.lattice_model_path = _resolve_model_path("LattifAI/Lattice-1")
     transcriber = create_transcriber(transcription_config=transcription_config)
 
-    print(colorful.cyan(f"🎤 Starting transcription with {transcriber.name}..."))
-    print(colorful.cyan(f"    Input: {input}"))
+    safe_print(colorful.cyan(f"🎤 Starting transcription with {transcriber.name}..."))
+    safe_print(colorful.cyan(f"    Input: {media_config.input_path}"))
 
     # Perform transcription
     if is_url and transcriber.supports_url:
         # Check if transcriber supports URL directly
-        print(colorful.cyan("    Transcribing from URL directly..."))
-        transcript = asyncio.run(transcriber.transcribe(input))
+        safe_print(colorful.cyan("    Transcribing from URL directly..."))
+        transcript = asyncio.run(transcriber.transcribe(media_config.input_path))
     else:
         if is_url:
             # Download media first, then transcribe
-            print(colorful.cyan("    Downloading media from URL..."))
+            safe_print(colorful.cyan("    Downloading media from URL..."))
             from lattifai.workflow.youtube import YouTubeDownloader
 
             downloader = YouTubeDownloader()
             input_path = asyncio.run(
                 downloader.download_media(
-                    url=input,
+                    url=media_config.input_path,
                     output_dir=str(output_path),
-                    media_format=media_format,
-                    force_overwrite=False,
+                    media_format=media_config.normalize_format(),
+                    force_overwrite=media_config.force_overwrite,
                 )
             )
-            print(colorful.cyan(f"    Media downloaded to: {input_path}"))
+            safe_print(colorful.cyan(f"    Media downloaded to: {input_path}"))
         else:
-            input_path = Path(str(input))
+            input_path = Path(media_config.input_path)
 
-        print(colorful.cyan("    Loading audio..."))
+        safe_print(colorful.cyan("    Loading audio..."))
         # For files, load audio first
         audio_loader = AudioLoader(device=transcription_config.device)
-        media_audio = audio_loader(input_path, channel_selector=channel_selector)
+        media_audio = audio_loader(
+            input_path,
+            channel_selector=media_config.channel_selector,
+            streaming_chunk_secs=media_config.streaming_chunk_secs,
+        )
         transcript = asyncio.run(transcriber.transcribe(media_audio))
 
     # Determine output caption path
@@ -153,14 +168,14 @@ def transcribe(
             final_output = output_path / f"youtube_LattifAI_{transcriber.name}.{output_format}"
         else:
             # For files, use input filename with suffix
-            final_output = Path(str(input)).with_suffix(".LattifAI.srt")
+            final_output = Path(media_config.input_path).with_suffix(".LattifAI.srt")
 
-    print(colorful.cyan(f"   Output: {final_output}"))
+    safe_print(colorful.cyan(f"   Output: {final_output}"))
 
     # Write output
     transcriber.write(transcript, final_output, encoding="utf-8", cache_audio_events=False)
 
-    print(colorful.green(f"🎉 Transcription completed: {final_output}"))
+    safe_print(colorful.green(f"🎉 Transcription completed: {final_output}"))
 
     return transcript
 

lattifai/cli/youtube.py

@@ -117,6 +117,7 @@ def youtube(
         force_overwrite=media_config.force_overwrite,
         split_sentence=caption_config.split_sentence,
         channel_selector=media_config.channel_selector,
+        streaming_chunk_secs=media_config.streaming_chunk_secs,
     )
 
 

lattifai/client.py

@@ -18,6 +18,7 @@ from lattifai.errors import (
     LatticeEncodingError,
 )
 from lattifai.mixin import LattifAIClientMixin
+from lattifai.utils import safe_print
 
 if TYPE_CHECKING:
     from lattifai.diarization import LattifAIDiarizer  # noqa: F401
@@ -91,6 +92,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
         input_caption_format: Optional[InputCaptionFormat] = None,
         split_sentence: Optional[bool] = None,
         channel_selector: Optional[str | int] = "average",
+        streaming_chunk_secs: Optional[float] = None,
     ) -> Caption:
         try:
             # Step 1: Get caption
@@ -100,6 +102,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
                 media_audio = self.audio_loader(
                     input_media,
                     channel_selector=channel_selector,
+                    streaming_chunk_secs=streaming_chunk_secs,
                 )
 
             if not input_caption:
@@ -113,7 +116,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
             alignment_strategy = self.aligner.config.strategy
 
             if alignment_strategy != "entire" or caption.transcription:
-                print(colorful.cyan(f"🔄 Using segmented alignment strategy: {alignment_strategy}"))
+                safe_print(colorful.cyan(f"🔄 Using segmented alignment strategy: {alignment_strategy}"))
 
                 if caption.supervisions and alignment_strategy == "transcription":
                     # raise NotImplementedError("Transcription-based alignment is not yet implemented.")
@@ -126,7 +129,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
                     if not caption.transcription:
                         import asyncio
 
-                        print(colorful.cyan("📝 Transcribing media for alignment..."))
+                        safe_print(colorful.cyan("📝 Transcribing media for alignment..."))
                         if output_caption_path:
                             transcript_file = (
                                 Path(str(output_caption_path)).parent
@@ -223,11 +226,11 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
                         continue
 
                     offset = round(start, 4)
-                    emission = self.aligner.emission(
-                        media_audio.tensor[
-                            :, int(start * media_audio.sampling_rate) : int(end * media_audio.sampling_rate)
-                        ]
-                    )
+                    # Extract audio slice
+                    audio_slice_ndarray = media_audio.ndarray[
+                        :, int(start * media_audio.sampling_rate) : int(end * media_audio.sampling_rate)
+                    ]
+                    emission = self.aligner.emission(audio_slice_ndarray)
 
                     # Align segment
                     _supervisions, _alignments = self.aligner.alignment(
@@ -259,7 +262,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
 
             # Step 5: Speaker diarization
             if self.diarization_config.enabled and self.diarizer:
-                print(colorful.cyan("🗣️  Performing speaker diarization..."))
+                safe_print(colorful.cyan("🗣️  Performing speaker diarization..."))
                 caption = self.speaker_diarization(
                     input_media=media_audio,
                     caption=caption,
@@ -308,7 +311,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
         if output_caption_path:
             diarization_file = Path(str(output_caption_path)).with_suffix(".SpkDiar")
             if diarization_file.exists():
-                print(colorful.cyan(f"Reading existing speaker diarization from {diarization_file}"))
+                safe_print(colorful.cyan(f"Reading existing speaker diarization from {diarization_file}"))
                 caption.read_speaker_diarization(diarization_file)
 
         diarization, alignments = self.diarizer.diarize_with_alignments(
@@ -433,12 +436,13 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
         split_sentence: Optional[bool] = None,
         use_transcription: bool = False,
         channel_selector: Optional[str | int] = "average",
+        streaming_chunk_secs: Optional[float] = None,
     ) -> Caption:
         # Prepare output directory and media format
         output_dir = self._prepare_youtube_output_dir(output_dir)
         media_format = self._determine_media_format(media_format)
 
-        print(colorful.cyan(f"🎬 Starting YouTube workflow for: {url}"))
+        safe_print(colorful.cyan(f"🎬 Starting YouTube workflow for: {url}"))
 
         # Step 1: Download media
         media_file = self._download_media_sync(url, output_dir, media_format, force_overwrite)
@@ -460,7 +464,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
         output_caption_path = self._generate_output_caption_path(output_caption_path, media_file, output_dir)
 
         # Step 4: Perform alignment
-        print(colorful.cyan("🔗 Performing forced alignment..."))
+        safe_print(colorful.cyan("🔗 Performing forced alignment..."))
 
         caption: Caption = self.alignment(
             input_media=media_audio,
@@ -468,6 +472,7 @@ class LattifAI(LattifAIClientMixin, SyncAPIClient):
             output_caption_path=output_caption_path,
             split_sentence=split_sentence,
             channel_selector=channel_selector,
+            streaming_chunk_secs=streaming_chunk_secs,
         )
 
         return caption

lattifai/config/alignment.py

@@ -18,8 +18,8 @@ class AlignmentConfig:
     """
 
     # Alignment configuration
-    model_name: str = "Lattifai/Lattice-1"
-    """Model identifier or path to local model directory (e.g., 'Lattifai/Lattice-1')."""
+    model_name: str = "LattifAI/Lattice-1"
+    """Model identifier or path to local model directory (e.g., 'LattifAI/Lattice-1')."""
 
     device: Literal["cpu", "cuda", "mps", "auto"] = "auto"
     """Computation device: 'cpu' for CPU, 'cuda' for NVIDIA GPU, 'mps' for Apple Silicon."""
@@ -58,6 +58,27 @@ class AlignmentConfig:
     Default: 4.0 seconds. Useful for detecting scene changes or natural breaks in content.
     """
 
+    # Beam search parameters for forced alignment
+    search_beam: int = 200
+    """Search beam size for beam search decoding. Larger values explore more hypotheses but are slower.
+    Default: 200. Typical range: 20-500.
+    """
+
+    output_beam: int = 80
+    """Output beam size for keeping top hypotheses. Should be smaller than search_beam.
+    Default: 80. Typical range: 10-200.
+    """
+
+    min_active_states: int = 400
+    """Minimum number of active states during decoding. Controls memory and search space.
+    Default: 400. Typical range: 30-1000.
+    """
+
+    max_active_states: int = 10000
+    """Maximum number of active states during decoding. Prevents excessive memory usage.
+    Default: 10000. Typical range: 1000-20000.
+    """
+
     client_wrapper: Optional["SyncAPIClient"] = field(default=None, repr=False)
     """Reference to the SyncAPIClient instance. Auto-set during client initialization."""
 

lattifai/config/caption.py

@@ -48,7 +48,7 @@ class CaptionConfig:
     include_speaker_in_text: bool = True
     """Preserve speaker labels in caption text content."""
 
-    normalize_text: bool = False
+    normalize_text: bool = True
     """Clean HTML entities and normalize whitespace in caption text."""
 
     split_sentence: bool = False

lattifai/config/media.py

@@ -52,12 +52,23 @@ class MediaConfig:
     sample_rate: Optional[int] = None
     """Audio sample rate in Hz (e.g., 16000, 44100)."""
 
-    channels: Optional[int] = None
-    """Number of audio channels (1=mono, 2=stereo)."""
-
     channel_selector: Optional[str | int] = "average"
     """Audio channel selection strategy: 'average', 'left', 'right', or channel index."""
 
+    # Audio Streaming Configuration
+    streaming_chunk_secs: Optional[float] = 600.0
+    """Duration in seconds of each audio chunk for streaming mode.
+    When set to a value (e.g., 600.0), enables streaming mode for processing very long audio files (>1 hour).
+    Audio is processed in chunks to keep memory usage low (<4GB peak), suitable for 20+ hour files.
+    When None, disables streaming and loads entire audio into memory.
+    Valid range: 1-1800 seconds (minimum 1 second, maximum 30 minutes).
+    Default: 600 seconds (10 minutes).
+    Recommended: Use 60 seconds or larger for optimal performance.
+    - Smaller chunks: Lower memory usage, more frequent I/O
+    - Larger chunks: Better alignment context, higher memory usage
+    Note: Streaming may add slight processing overhead but enables handling arbitrarily long files.
+    """
+
     # Output / download configuration
     output_dir: Path = field(default_factory=lambda: Path.cwd())
     """Directory for output files (default: current working directory)."""
@@ -87,12 +98,21 @@ class MediaConfig:
         self._normalize_media_format()
         self._process_input_path()
         self._process_output_path()
+        self._validate_streaming_config()
 
     def _setup_output_directory(self) -> None:
         """Ensure output directory exists and is valid."""
         resolved_output_dir = self._ensure_dir(self.output_dir)
         self.output_dir = resolved_output_dir
 
+    def _validate_streaming_config(self) -> None:
+        """Validate streaming configuration parameters."""
+        if self.streaming_chunk_secs is not None:
+            if not 1.0 <= self.streaming_chunk_secs <= 1800.0:
+                raise ValueError(
+                    f"streaming_chunk_secs must be between 1 and 1800 seconds (1 second to 30 minutes), got {self.streaming_chunk_secs}. Recommended: 60 seconds or larger."
+                )
+
     def _validate_default_formats(self) -> None:
         """Validate default audio and video formats."""
         self.default_audio_format = self._normalize_format(self.default_audio_format)