Audio & Voice¶
Modern speech systems span three core areas: recognizing spoken language (ASR), synthesizing voices (TTS), and converting or cloning speaker identity. Articles here cover current model families (Whisper, Qwen3-ASR, NVIDIA Canary, F5-TTS, CosyVoice), latency budgets for real-time voice agents, fine-tuning infrastructure on rented GPUs, and multilingual deployment trade-offs. Each piece is a dense reference: architecture diagrams, commands, benchmarks, and integration gotchas — not tutorials.
Speech & Recognition¶
- speech recognition - ASR models, transcription, pronunciation assessment
Text-to-Speech¶
- tts models - TTS model comparison, latency benchmarks, multilingual support
- voice cloning - Voice cloning, voice mixing, naturalness benchmarks
- voice conversion - Voice conversion techniques and pipelines
- audio generation - Audio generation models and workflows
- audio flamingo - Audio Flamingo 3, Music Flamingo, and AF-Next understanding artifacts
- ace step 1 5 - ACE-Step 1.5 base/SFT/turbo and XL hardware bounds
Voice Applications¶
- voice agent pipelines - Voice agent pipelines and frameworks for real-time applications
- podcast processing - Podcast processing, transcription, and analysis
Additional References¶
- asr stt compression - KV cache compression methods for ASR/TTS inference and LLM context in 2026: TriAttention
- audio omni unified model - Single model for audio understanding, generation, and editing via frozen LLM reasoning + trainable
- lemas tts and speech editing - LEMAS open-source multilingual TTS and word-level speech editing models - architecture
- tts fine tuning infrastructure - GPU rental platform comparison and deployment patterns for fine-tuning and serving 2B-4B TTS models
- voice design - Creating unique synthetic voices from text descriptions, voice morphing, naturalness benchmarks