AI Voice Guides

Practical guides to AI voice cloning, text to speech, recording, pronunciation, and creating better voiceovers.

Written byRandy Wake

Connect with me

Latest voice guides

MOSS-VoiceGenerator voice-design diagram connecting a descriptive brief to timbre, style, and emotion and then generated speech for listening review.

Voice Design Research

Oct 29, 20266 min read

MOSS-VoiceGenerator Research: Designing Voices in Words

Learn how MOSS-VoiceGenerator turns descriptions into voices, where its research finds limits, and how to build a repeatable voice-design audition.

Randy Wake

Chatterbox multilingual diagram showing speech generation, watermark embedding, and a separate verification step.

Multilingual Speech Research

Oct 26, 20266 min read

Chatterbox v3 Research: Languages and Audio Watermarks

Review Chatterbox Multilingual v3's language coverage, PerTh watermarking, and evaluation limits, with a practical plan for checking your exported audio.

Randy Wake

Pocket TTS local-evaluation diagram showing script, a compact model on a laptop, and audio with checks for startup, playback, and clarity.

Speech Model Research

Oct 23, 20266 min read

Pocket TTS Research: Small Models for Local Speech

Understand Pocket TTS's continuous-audio research, CPU-focused design, language variants, and checks for local narration and responsive audio playback.

Randy Wake

Dia2 diagram with two speakers taking turns as text arrives and audio is generated in incremental chunks.

Streaming Speech Research

Oct 20, 20266 min read

Dia2 Research: Streaming Dialogue and Speaker Context

Explore Dia2's streamed text input, two-speaker conditioning, and English dialogue limits, with a practical test for turn timing and consistent voice identity.

Randy Wake

LongCat-AudioDiT diagram showing waveform-latent diffusion and separate reconstruction and speech quality checks.

Speech Model Research

Oct 17, 20266 min read

LongCat-AudioDiT Research: Speech in Waveform Latents

Understand LongCat-AudioDiT's waveform-latent approach, its reconstruction tradeoff, and a practical listening test for faithful, intelligible voice cloning.

Randy Wake

Confucius4-TTS conceptual pathway where an audio prompt without a transcript goes through a speaker encoder, and both its information and target text enter a text-to-semantic model before flow matching produces speech.

Multilingual Speech Research

Oct 14, 20266 min read

Confucius4-TTS Research: Cloning Without Transcripts

Learn how Confucius4-TTS uses speaker conditioning for transcript-free cloning, what its reported results cover, and how to evaluate cross-language identity.

Randy Wake

IndexTTS 2.5 conceptual evaluation loop showing text, a voice reference, and an expression cue producing speech that is checked separately for word accuracy, speaker identity, and expressive fit by a listening panel.

Speech Model Research

Oct 11, 20266 min read

IndexTTS 2.5 Research: Expression, Speed, and Control

Explore IndexTTS 2.5's multilingual research, emotion and pronunciation controls, and release differences with a plan for evaluating expressive speech.

Randy Wake

Voxtral TTS conceptual pipeline showing a text and voice prompt entering an autoregressive planner, producing semantic tokens, then a continuous acoustic generator and speech.

Streaming Speech Research

Oct 8, 20266 min read

Voxtral TTS Research: Hybrid Speech for Voice Agents

Understand Voxtral TTS's hybrid speech architecture, streaming evaluation, voice adaptation, and model license before planning a voice-agent experiment.

Randy Wake

X-Voice conceptual diagram separating reference audio and its language from target text and target language so speaker traits can lead to target speech without a reference transcript.

Multilingual Speech Research

Oct 5, 20266 min read

X-Voice Research: Cross-Lingual Cloning Without Reference Text

Examine X-Voice's transcript-free reference method, multilingual training choices, and practical tests for pronunciation, accent, and speaker identity.

Randy Wake

OmniVoice conceptual multilingual pipeline from text and a language tag through a diffusion language model and acoustic tokens to speech, paired with a reminder that language coverage needs separate quality evaluation.

Multilingual Speech Research

Oct 2, 20266 min read

OmniVoice Research: What 600-Language TTS Means

Explore OmniVoice's multilingual diffusion design, distinguish language coverage from evaluated quality, and build a useful pilot for underserved audiences.

Randy Wake

MOSS-TTS diagram showing 24 kHz audio compressed into 12.5 fps tokens and splitting into separate MOSS-TTS long-context and Local Transformer frame-local alternatives, each with its own speech path.

Speech Model Research

Sep 29, 20266 min read

MOSS-TTS Research: Why the Local Transformer Matters

Compare MOSS-TTS's delay and local-transformer designs, understand checkpoint differences, and plan meaningful tests of speaker consistency and control.

Randy Wake

VoxCPM2 conceptual diagram where voice design, controlled cloning, and continuation are alternative inputs to continuous audio latents, local diffusion, and a 48 kHz waveform.

Speech Model Research

Sep 26, 20266 min read

VoxCPM2 Research: Continuous Speech and Voice Control

Understand VoxCPM2's continuous audio latents, voice design and cloning modes, and the evaluation questions behind its multilingual speech research.

Randy Wake