Google Cloud

Lyria 3.5 (lyria-3.5) is Google's newest music model, GA in the Gemini API since September 3, 2026, generating full-length songs with vocals and timed lyrics from text or image prompts as 44.1 kHz stereo. It replaces Lyria 3 Pro.

Lyria 3 Clip (lyria-3-clip-preview), a March 2026 preview, always produces a 30-second MP3 clip, vocal or instrumental, and returns the lyrics and song structure with it. Google's docs pair it with Lyria 3.5 for quick prompt drafts before full songs.

Chirp 3 (chirp_3) is Google Cloud's newest Speech-to-Text model, GA in the V2 API since October 2025. It adds automatic language detection, batch speaker diarization and a denoiser, with 29 languages GA and dozens more in preview.

Chirp 3: HD is Cloud Text-to-Speech's generative voice family, GA since April 2025, with 30 voices across about 53 locales, streaming and batch synthesis, and pace, pause and pronunciation controls. Gemini-TTS models sit above it in Google Cloud's lineup.

Telephony (telephony) is a Speech-to-Text V2 model tuned for phone-call audio, typically 8 kHz, launched in November 2023 as the latest version of the phone_call model. It keeps speaker diarization; Chirp 3 is Google's newer general-purpose recognizer.

Studio voices are Cloud Text-to-Speech voices introduced in February 2023 for narration and news reading, with 11 voices across six English, Spanish, French and German locales. They accept most SSML; Chirp 3: HD is the newer family.

Neural2 is a Cloud Text-to-Speech voice tier introduced in June 2022 that applies Custom Voice technology without training, with about 60 voices across 18 locales and SSML support. Still GA, it predates Chirp 3: HD and Gemini-TTS.

Standard voices are Cloud Text-to-Speech's original parametric, vocoder-based tier, offered since the March 2018 launch. They are the lowest-cost option, with full SSML and the broadest locale list (about 59), but the oldest technology in Google's lineup.

WaveNet voices use DeepMind's WaveNet raw-audio model and have been in Cloud Text-to-Speech since its March 2018 launch, now about 160 voices across 46 locales with SSML. A step above Standard, they predate Neural2, Chirp 3: HD and Gemini-TTS.

Chirp 2 (chirp_2) is a Speech-to-Text V2 model built on Google's Universal Speech Model, offering streaming, sync and batch recognition with speech translation, denoising and word timestamps but no diarization. Chirp 3, GA since October 2025, is the newer generation.

Polyglot is a small Cloud Text-to-Speech voice set whose voices can speak more than one language, such as en-US-Polyglot-1, with six voices across English, Spanish, French and German locales. It runs only on multi-region endpoints.