speech▌

Transcribe audio and video to text with speaker identification, word-level timestamps, and 90+ language support. \n \n Two models available: scribe_v2 for batch transcription with high accuracy, and scribe_v2_realtime for live transcription with ~150ms latency \n Speaker diarization labels each word with speaker ID; keyterm prompting helps recognize domain-specific vocabulary and proper nouns \n Word-level timestamps include type classification (word, spacing, audio event) for precise timing and

Convert text to natural speech via inference.sh CLI.

Generate speech audio from text using HeyGen's Starfish TTS model with voice, speed, and pitch control. \n \n List available TTS voices by language and gender, then generate audio files with customizable speed (0.5–1.5) and pitch (−50 to 50) \n Supports multilingual voices with locale selection (e.g., pt-BR ) and SSML-style break tags for pauses within text \n Returns audio URL, duration, request ID, and word-level timestamps for caption syncing or timed overlays \n Requires HEYGEN_API_KEY envir

Transcribe live and pre-recorded audio to text using Apple's Speech framework. Covers SFSpeechRecognizer (iOS 10+) and the new SpeechAnalyzer API (iOS 26+).

Routes NVIDIA Nemotron Speech (Riva) NIM tasks — deploys, runs, and tests ASR, TTS, and NMT NIMs on build.nvidia.com or self-hosted.

File Organization: Split structure. See references/ for detailed implementations.

Transcribe audio to text via inference.sh CLI.

Text-to-speech generation for narration, voiceovers, IVR prompts, and accessibility reads via OpenAI Audio API. \n \n Supports single clips and batch processing; defaults to gpt-4o-mini-tts-2025-12-15 with built-in voices (cedar, marin, and others) \n Includes instruction augmentation for voice affect, tone, pacing, emotion, and emphasis; instructions supported only on GPT-4o mini TTS models \n Enforces 4096-character input limit per request and 50 requests/minute rate cap; splits longer text in

Natural speech synthesis from text across 70+ languages with multiple quality and latency models. \n \n Six models available ranging from highest-quality eleven_v3 to ultra-low-latency eleven_flash_v2_5 (~75ms), with language and speed tradeoffs documented \n Supports 13+ output formats including MP3, PCM, WAV, Opus, and telephony codecs (μ-law, A-law) for web, streaming, and real-time applications \n Fine-tune voice characteristics via stability, similarity boost, style, speaker boost, and spee

Multiple text-to-speech models via inference.sh CLI for voiceovers, podcasts, and accessibility. \n \n Six models available: ElevenLabs (premium, 22+ voices, 32 languages), DIA TTS (conversational), Kokoro TTS (fast), Chatterbox, Higgs Audio (emotional control), and VibeVoice (long-form podcasts) \n Core capabilities include basic speech synthesis, expressive speech with emotion control, and conversational dialogue generation \n Easily combine with video tools like OmniHuman to create talking he