If you're building a voice agent, transcription is the first domino. Get it wrong and every downstream stage — your LLM, your logic, your text-to-speech — inherits garbage. The Google Speech-to-Text API is one of the most battle-tested options for turning audio into text, but the reference docs read like a compliance manual and every "review" is written by a competitor selling their own model.
This is the vendor-neutral version. We'll cover what the API actually does, what it costs in 2026, how to ship working code, and — the part nobody tells you — where raw speech-to-text stops being enough for a real-time phone agent.
What is the Google Speech-to-Text API?
Google Speech-to-Text (STT) is a cloud API that converts audio into text using Google's automatic speech recognition (ASR) models. It exposes three recognition modes, and picking the right one is the single biggest architecture decision you'll make:
- Synchronous — send a short audio clip (≤ 60 seconds), block, get the full transcript back in one response. Simplest to code; useless for live calls.
- Asynchronous / batch (
LongRunningRecognize) — upload long audio (up to ~480 minutes) to Cloud Storage, poll for a result. Right for call-recording pipelines, voicemail, podcast transcription. - Streaming (
StreamingRecognize) — a bidirectional gRPC stream where you push audio chunks and receive interim and final results as the caller talks. This is the mode any real-time voice agent lives or dies on.
On the model side, the Chirp family (Google's universal speech models, chirp and chirp_2 in the v2 API) is the 2026 default for accuracy and multilingual coverage. Older latest_long / latest_short and phone-tuned models still exist in v1 and matter when you need a specific feature (like certain telephony enhancements) that a newer model hasn't shipped yet. Always match the model to the audio: 8kHz mono telephony audio through a model tuned for 16kHz studio audio is a self-inflicted accuracy wound.
Google STT API pricing in 2026
Pricing is per-minute of audio processed, billed in rounded increments, with tiers that differ by model and features. Rather than quote numbers that drift, here's how to reason about the bill — and where it bites at scale.
Verify before you commit: always price against the live Cloud Speech-to-Text pricing page. The figures below describe the structure of billing in 2026, not a locked quote.
- Free tier: Google has historically offered a monthly free allotment (on the order of ~60 minutes) — enough to prototype, nowhere near enough to run.
- Standard vs enhanced/premium models: newer/enhanced models bill at a higher per-minute rate than legacy standard recognition. Chirp-class models sit at the premium end.
- Features stack on top: speaker diarization, word-level confidence, and some model options can change the effective rate or add processing.
- Logging discount: opting into data logging (letting Google use your audio to improve models) historically unlocked a lower rate. Read the data-governance implications before trading transcript data for a discount — for regulated workloads that trade is usually a non-starter.
The hidden cost at scale is rounding + always-on streaming. Batch jobs round per request. Streaming bills for the duration the stream is open — including the silence while your agent is thinking or speaking if you leave the mic hot. At 10,000 concurrent calls, a few wasted seconds of open stream per turn compounds into real money. Close the stream on end-of-utterance; don't hold it open for the whole call.
Quickstart: authenticate + transcribe audio
Two ways in. Authenticate first: create a service account in Google Cloud, grant it the Speech-to-Text role, download the JSON key, and point GOOGLE_APPLICATION_CREDENTIALS at it.
