Google Cloud Chirp 3: model profile

Overview

Chirp 3 is a brand umbrella spanning automatic speech recognition, HD text-to-speech, and custom voice creation, each with different APIs, limits, prices, and regional footprints. Google describes Chirp 3: Transcription as the latest generation of its multilingual ASR-specific generative models, available only in Speech-to-Text API V2 under the model identifier chirp_3. Google's product page states that Speech-to-Text can use Chirp 3, trained on millions of hours of audio and billions of text sentences, to improve recognition across more languages and accents. Google positions Chirp 3 for multilingual transcription and for lower latency synthetic speech in agents, support operations, and media workflows.

The transcription side and the synthesis side are separate service surfaces. Deployment characteristics, compliance controls, and operational limits differ between the STT and TTS products despite the shared branding.

Capabilities and features

Speech-to-Text (chirp_3)

Chirp 3 transcription maps to three API modes: Speech.Recognize for audio shorter than one minute, Speech.BatchRecognize for long audio (documented as generally suited for 1 minute to 1 hour, though quotas allow longer), and Speech.StreamingRecognize for streaming and real-time audio. A BatchRecognize request can contain up to 15 files, each up to 8 hours long; long-audio batch requests accept input only as a Cloud Storage URI.

Documented features:

• Automatic punctuation
• Automatic capitalization
• Utterance-level timestamps in streaming
• Speaker diarization in batch
• Speech adaptation for vocabulary biasing
• Language-agnostic audio transcription
• Custom prompt feature (preview)

Documented caveats: word-level timestamps are supported only in Recognize and BatchRecognize and can degrade transcription quality; word-level confidence scores are returned but are not confidence scores in the conventional sense.

For live systems, Chirp 3 exposes endpointing_sensitivity levels (standard, short, supershort) that trade accuracy against latency. Google's documentation warns that increasing sensitivity for faster bot responses can cause the system to cut off users who pause briefly while speaking. Speech-to-Text V2 can emit voice activity begin and end events and can auto-close streams based on speech begin or speech end timeouts; those timeouts must be greater than 500 ms and less than 60 s.

Text-to-Speech (Chirp 3 HD voices)

Google's documentation says Chirp 3 HD voices are powered by generative models, are suitable for conversational agents, cover 30 distinct styles across many languages, and support low-latency real-time communication through text streaming. Streaming Cloud TTS is only compatible with Chirp 3 HD voices.

Output formats: LINEAR16 is the default; ALAW, MULAW, OGG_OPUS, and PCM are supported for streaming; ALAW, MULAW, MP3, OGG_OPUS, and PCM are supported for batch.

Control features:

• Pace control from 0.25x to 2.0x
• Pause control (experimental)
• Custom pronunciations using IPA or X-SAMPA
• SSML support in preview for synchronous Chirp 3 HD requests; those SSML tags are not currently supported for streaming requests
• Asynchronous long-form synthesis of up to 1 million bytes of input

Instant Custom Voice

Instant Custom Voice creates personalized voice models and synthesizes speech through the Cloud TTS API for both streaming and long-form text. Access is allowlist-only. Google requires both a consent recording and a reference audio recording, each up to 10 seconds, with the consent statement using Google's required script rather than a customer-written alternative. The generated artifact is a voice cloning key, stored on the client side and reusable across multiple clients or devices. Google's April 2025 blog states the feature enables realistic custom voices from 10 seconds of input, targets use cases such as personalized call centers, accessible content, and brand voices, and that Google uses built-in safety features and a diligence process to verify voice-use permissions.

Instant Custom Voice supports language transfer from en-US to several other locales, including de-DE, es-US, es-ES, fr-CA, fr-FR, and pt-BR.

Language support

Transcription: Chirp 3 launched publicly with 85+ languages and locales.

Diarization: the supported subset is smaller. The Chirp 3 model page lists diarization support for specific languages including en-US, en-GB, en-IN, fr-FR, fr-CA, es-ES, es-US, de-DE, hi-IN, it-IT, ja-JP, ko-KR, pt-BR, and Simplified Chinese.

Synthesis: Google's April 2025 blog said Chirp 3 HD voices offered natural speech in over 35 languages with eight speaker options, while release notes described the GA catalog as 8 speakers and 31 locales. The two figures count languages and locale variants respectively.

Performance and benchmarks

Vendor-reported: Google does not publish WER figures for Chirp 3 in the source material. The 2023 USM research that preceded Chirp benchmarked against Whisper with multilingual gains.

Third-party evaluation: a 2026 academic evaluation on noisy Dutch semi-spontaneous speech found Google Chirp 3 achieved the best average WER among eight tested ASR systems at 11.2%, compared with Whisper-large-v3 at 15.8%. A separate 2026 study on spoken U.S. street names found that leading systems from Google, OpenAI, Deepgram, and Microsoft produced a very high average error rate on this named-entity task.

Latency and throughput

Google publishes developer controls rather than a public millisecond latency benchmark.

On STT, endpointing sensitivity controls how quickly the model finalizes utterances, and voice activity events can fire before the corresponding transcription result. On TTS, Chirp 3 HD supports text streaming, letting applications send text incrementally and receive audio incrementally.

Deployment and integrations

Chirp 3 transcription is available only in Speech-to-Text API V2 under the model identifier chirp_3. The Chirp 3 model page lists us and eu multi-regions as GA for the transcription model; Google's STT V2 migration material says the broader platform supports regionalized invocation in locations such as Belgium and Singapore, so model-specific regional availability must be checked separately from platform-wide regionalization.

Chirp 3 HD voice documentation lists regional availability in global, us, eu, asia-southeast1, europe-west2, and asia-northeast1.

Setup requires a Google Cloud project with billing and the relevant APIs enabled; enabling Speech-to-Text or Text-to-Speech requires the Service Usage Admin role for the person enabling the APIs. For STT V2, Google recommends recognizers, which are reusable stored recognition configurations for grouping traffic and standardizing model, language, and feature settings. BatchRecognize expects source audio in Cloud Storage and returns a long-running operation whose output can be returned inline or written back to Cloud Storage. Long-form TTS requires the output bucket to grant Storage Object Creator and Storage Object Viewer roles.

Compliance controls: Speech-to-Text supports IAM and audit logging; Text-to-Speech offers regional endpoints, and Google says using us or eu regional endpoints keeps TTS data at rest and in use within those geographic boundaries. Google states that Cloud TTS does not log customer text or audio data. In Vertex AI generative media launch materials for Chirp 3, Google said it does not use customer data to train its models and that customer data is processed according to customer instructions.

Pricing

Cost scenarios given in the source:

Comparative pricing reported in the source: OpenAI lists about $0.006/minute and $0.003/minute for its two request-based transcription models and $0.017/minute for GPT-Realtime-Whisper. ElevenLabs lists $0.05 per 1K characters for Flash/Turbo TTS, $0.10 per 1K for Multilingual v2/v3, $0.22/hour for Scribe v1/v2 STT, and $0.39/hour for realtime STT. Deepgram lists Aura-2 TTS at $0.030 per 1K characters. Microsoft pricing snippets show standard transcription at about $1/hour and standard/neural TTS at about $15 per 1M characters. AWS Polly spans $16 per 1M characters for neural voices and $30 per 1M for generative voices.

Development and ownership

Chirp 3 is developed by Google Cloud. Its lineage begins with Google Research's Universal Speech Model (USM) work, described in March 2023 as part of Google's 1,000 Languages Initiative. Yu Zhang and James Qin presented USM in the research blog; the corresponding paper lists 27 authors and describes a 2B-parameter multilingual ASR system trained on 12 million hours of speech and 28 billion sentences of text spanning 300+ languages.

Google Cloud productized that work in May 2023, when Cloud Speech product manager Calum Barnes introduced Chirp as a foundation model for the Google Cloud Speech API, described as a 2B-parameter speech model built through self-supervised training on millions of hours of audio and 28 billion sentences across 100+ languages.

Release history