Google Cloud Speech-to-Text

Google Cloud Speech-to-Text is Google's managed automatic speech recognition (ASR) API, part of Google Cloud's AI/ML portfolio. It exists in two generations: the legacy v1 API and the modern v2 API, the latter built around Google's Chirp family of foundation models (Chirp, Chirp 2, and the current Chirp 3), which are large self-supervised multilingual models trained on millions of hours of audio across 100+ languages. The product targets developers and enterprises already on Google Cloud who need transcription embedded in IAM, VPC-SC, CMEK and audit-logging governance, call-center analytics, media captioning, voice agents, and meeting transcription are the dominant use cases. It supports synchronous, streaming (real-time), and asynchronous batch recognition, plus speaker diarization, automatic punctuation, word-level timestamps, model adaptation/custom vocabulary, and a built-in denoiser.

Where it wins is breadth and platform integration: 85–125+ languages and locales depending on model, deep GCP-native security/compliance, a generous lower v2 price point ($0.016/min standard, dropping to ~$0.004/min at volume and a Dynamic Batch tier ~75% cheaper for non-urgent jobs), and a mature multi-language SDK surface. Where it loses is accuracy on hard real-world audio and developer experience. Independent benchmarking (Voicewriter.io, cited by Deepgram) put Google at a 13.1% word error rate on noisy/accented/technical audio, materially worse than specialist competitors like Deepgram Nova (~7.6%) and roughly comparable-to-worse than OpenAI Whisper (~10.6%), so its leadership in clean conversational English does not always carry into production conditions.

The other recurring friction is operational: the v1/v2 split with differing features, pricing and regional availability confuses integrators; speaker diarization and some add-ons carry extra cost that is not transparently surfaced on the public pricing page; and billing can "double" versus the headline rate once GCP overhead and add-ons are counted. For teams already standardized on Google Cloud it is a safe, well-governed default with a strong SLA and free monthly tier; teams optimizing purely for transcription accuracy, simpler pricing, or DX increasingly evaluate Deepgram, AssemblyAI, or Whisper-based stacks alongside it.