AllVoiceLab

Generates lifelike AI-synthesized speech from text input across 30+ languages using the proprietary MaskGCT model, which enables emotionally expressive and tonally varied speech synthesis. The system supports multiple speaking styles and tones per language, allowing developers to control prosody and emotional delivery without manual voice recording or post-processing. Integration occurs via MCP tool invocation with text input and audio file output.

Clones a speaker's voice from a short audio sample (claimed to work in seconds) by extracting and encoding speaker characteristics including pitch, rhythm, and emotional tone, then applying those characteristics to new text-to-speech synthesis. The system operates as a write-once operation that produces new audio artifacts with the cloned voice characteristics applied. Implementation details of the speaker encoding mechanism are proprietary and undocumented.

Transforms input audio by modifying voice characteristics (pitch, timbre, accent) in real-time or near-real-time without requiring speaker-specific model training or fine-tuning. The system accepts audio input and applies voice transformation rules or learned transformations to produce modified audio output. Specific transformation parameters and the underlying voice encoding mechanism are proprietary.

Extracts clean vocal tracks from mixed audio by applying source separation techniques to isolate voice from background music, noise, and other non-vocal elements. The system accepts audio input and produces isolated vocal and instrumental tracks as separate output files. Implementation uses neural source separation but specific model architecture and training data are proprietary.

Automates the complete video dubbing workflow by accepting video input, extracting dialogue, translating to target language(s), synthesizing new audio in target language with voice cloning or TTS, and re-synchronizing audio with video. The system orchestrates multiple sub-operations (transcription, translation, TTS, audio mixing, video re-encoding) into a single end-to-end pipeline. Specific translation engine and synchronization algorithm are undocumented.

Analyzes video input to detect, transcribe, and time-align subtitles with >98% accuracy claimed. The system performs optical character recognition (OCR) on video frames to identify hardcoded subtitles, transcribes their text content, and aligns timing with video timeline. Output includes subtitle file (SRT, VTT, or similar) with timing metadata. This is a read-only analysis operation that does not modify the video.

Removes hardcoded (burned-in) subtitles from video by detecting subtitle regions and reconstructing background content using inpainting or content-aware fill techniques. The system accepts video input, identifies subtitle bounding boxes and timing, and generates new video frames with subtitles removed and backgrounds reconstructed. Output is a modified video file without visible subtitles. This is a write-once operation that produces a new video artifact.

Exposes AllVoiceLab voice and video processing capabilities as an MCP (Model Context Protocol) server, enabling AI agents and LLM-based applications to invoke voice synthesis, cloning, isolation, and video dubbing operations as tool calls within agent reasoning loops. The MCP server abstracts underlying API complexity and provides standardized tool schemas for agent integration. Transport mechanism (stdio, SSE, HTTP) and authentication flow are undocumented.

Supports batch processing of multiple audio or video files through asynchronous job submission and status polling. The system accepts batch input (multiple files or file lists), queues processing jobs, and provides job status tracking and result retrieval via polling or webhooks. Specific job queue implementation, concurrency limits, and result storage mechanism are undocumented.