Chirp 3: model profile

Overview

Google uses the name "Chirp 3" in official documentation for three related cloud services: Chirp 3: HD voices for text-to-speech, Chirp 3: Instant Custom Voice for voice cloning, and Chirp 3: Transcription for speech-to-text. The family launched in phases during 2025 rather than on a single date: Chirp 3: HD voices reached general availability on April 2, 2025, Instant Custom Voice was announced as GA through an allowlist on April 9, 2025, and Chirp 3: Transcription entered private preview on April 11, 2025, public preview on August 29, 2025, and GA on October 13, 2025.

Chirp 3 is a cloud-managed speech stack, not a firmware-based device. There is no public firmware image or consumer software versioning; capability changes are surfaced through release notes, region rollouts, model identifiers, and documentation updates.

The name "Chirp 3" is ambiguous with a hardware product. The source resolves the ambiguity as follows.

Capabilities and features

The family comprises three variants.

For STT, Chirp 3 supports speaker diarization, automatic punctuation, automatic capitalization, speech adaptation with up to 1,000 phrases, a custom prompt feature in Preview, and a built-in denoiser that can reduce music, rain, and street noise but not background human voices.

For TTS, text streaming is exclusive to Chirp 3 HD voices in Google's Cloud TTS stack. Limited SSML support was added on October 17, 2025, covering the phoneme, p, s, sub, and say-as tags.

For custom voice, Google requires a spoken consent statement, recommends clean 10-second recordings, stores the resulting voice-cloning key client-side, and permits 10 new keys per minute per project with no stated absolute limit on total keys.

Language support

• Chirp 3: Transcription documentation lists 29 GA transcription locales, many additional Preview locales, and 14 diarization locales, plus language-agnostic transcription.
• At public preview (August 29, 2025), chirp_3 was announced with 85+ languages/locales in preview coverage.
• Chirp 3: HD voices documentation lists 30 named voices and 53 supported languages/locales.
• Instant Custom Voice added ja-JP on June 18, 2025, pushing support to more than 30 locales, and supports multilingual transfer from en-US to six listed locales.
• In November and December 2025, STT preview regions expanded and TTS added a wide set of European languages, then Punjabi and Cantonese in preview.

Performance and benchmarks

Vendor-reported: Google's 2023 launch materials for the original Chirp claimed 98% English recognition accuracy and 300% relative improvement in some low-resource languages. For Chirp 3, Google's launch materials emphasize feature breadth, speed improvements, diarization, language detection, and voice realism, and do not provide a public benchmark sheet of equal granularity.

Third-party evaluation: Artificial Analysis' streaming STT benchmark found that Chirp 3 Streaming led partial-transcript performance on VoxPopuli at 2.2% WER, while noting that no single model led across all tested datasets. Artificial Analysis' selected-voice TTS leaderboard snapshot placed Chirp 3: HD at Elo 1,056 and $30.0 per 1M characters, below Azure HD 2.5 at Elo 1,127 and Eleven v3 at Elo 1,179.

User sentiment reported by the source: G2 reviewers of Google Cloud Speech-to-Text praise ease of use, speed, and meeting-transcription productivity; recurring negatives include cost sensitivity and the need for manual correction when accuracy is not perfect. For Google Cloud Text-to-Speech, review summaries emphasize natural voice quality and simple API integration, while some users describe output as robotic in some scenarios or languages and some complain about pricing opacity or cost escalation. Google developer forum threads surfaced UI confusion, long-audio latency regressions, markup/SSML limitations, locale-specific pronunciation bugs, and allowlist/access friction.

The source compares Chirp 3 against alternative platforms as follows.

Latency and throughput

Specific latency figures for Chirp 3 are not publicly disclosed in the source set. The source records the following latency-related facts:

• Text streaming is exclusive to Chirp 3 HD voices in Google's Cloud TTS stack, which the source describes as relevant for low-latency voice agents.
• The chirp_3 public preview announcement (August 29, 2025) included improved speed/accuracy messaging.
• Adya reported localization across 20+ Indian languages with low latency using Chirp 3.
• A September 2025 forum thread reported long-audio jobs stalling or taking much longer than before; no official public postmortem or universal fix was captured in the source set.
• Instant Custom Voice permits 10 new voice cloning keys per minute per project with no stated absolute limit on total keys.

Deployment and integrations

Chirp 3 is sold as a Google Cloud API service, not a boxed consumer product. Consumption happens through the Cloud TTS API, Speech-to-Text V2 API, Vertex AI Studio, and Google's console and notebook ecosystem.

Region support differs by sub-product: Chirp 3 HD lists six GA endpoints (global, us, eu, asia-southeast1, europe-west2, asia-northeast1), while Chirp 3 Transcription documents GA in us and eu multi-regions with release-note preview expansions into additional regions in late 2025. Instant Custom Voice lists region availability beyond the initial TTS endpoints.

Chirp 3: Transcription is available only in Speech-to-Text V2 and supports the StreamingRecognize, Recognize, and BatchRecognize methods.

Official support paths include reference documentation, release notes with RSS support, Vertex AI Studio and console entry points, Colab and GitHub notebooks, community forums, Cloud support, system status, and sales-led access for allowlisted features such as Instant Custom Voice.

Google describes Gemini-TTS as the latest evolution of Cloud TTS, with broader prompt-based control and native multi-speaker options while reusing voice identities similar to Chirp 3 HD.

Pricing

The source notes that Google's public pricing nomenclature lags the model nomenclature: the STT pricing page references "chirp (Speech-to-Text V2 only)" while the current model docs and release notes use chirp_3. The source's reading is that Chirp-family Speech-to-Text V2 pricing applies, but the public pricing page is not as current or precise as the model documentation.

Development and ownership

Chirp 3 is developed and operated by Google as part of Google Cloud. The Chirp family predates Chirp 3: Google introduced Chirp as a speech foundation model in 2023 and described that generation as a 2B-parameter speech model delivering 98% English speech recognition accuracy and large relative gains in some low-resource languages. Current Chirp 3 materials emphasize product capabilities, rollout stages, and API behavior rather than architectural disclosures such as parameter count.

Release history