Transcribe long audio stored in Cloud Storage with Google Cloud Speech-to-Text v2 BatchRecognize and retrieve transcripts

Documented steps

1. Batch recognition (BatchRecognize) is the v2 path for audio longer than 60 seconds, up to a documented upper limit of 480 minutes (8 hours) per file. For shorter audio use Recognize instead.
2. Upload audio to a Cloud Storage bucket and note the gs://BUCKET/OBJECT URI. The caller's service account needs read access to that bucket and write access to any output bucket.
3. Pick a recognizer. To use the implicit default without creating one, use the literal recognizer ID '_': projects/{PROJECT_ID}/locations/global/recognizers/_ .
4. Optional: create a custom recognizer with CreateRecognizer under parent=projects/{PROJECT_ID}/locations/{LOCATION}, setting default_recognition_config.language_codes (e.g. ["en-US"]) and default_recognition_config.model (e.g. 'long', 'latest_long', 'chirp_2'). Creating a recognizer requires the Cloud Speech Administrator IAM role.
5. POST to https://speech.googleapis.com/v2/{recognizer}:batchRecognize with Authorization: Bearer <access token> (e.g. gcloud auth print-access-token) and Content-Type: application/json.
6. In the body set config (RecognitionConfig) with autoDecodingConfig for common container formats, languageCodes, and model. Use configMask to declare which config fields override the recognizer defaults — omitting configMask means recognizer defaults win for unset fields.
7. Set files[] to BatchRecognizeFileMetadata objects, each with uri = gs://BUCKET/OBJECT.
8. Choose an output mode: recognitionOutputConfig.gcsOutputConfig.uri = gs://BUCKET/prefix writes one JSON transcript per input file; recognitionOutputConfig.inlineResponseConfig returns results inline in the operation and is only practical for small jobs.
9. Optionally set processingStrategy = DYNAMIC_BATCHING, documented as lower cost in exchange for higher latency and available only for batch recognition.
10. The call returns a long-running Operation. Poll GET https://speech.googleapis.com/v2/{operation_name} (or operation.result(timeout=...) in a client library) until done is true.
11. Read results: with gcsOutputConfig, fetch the per-file BatchRecognizeResults JSON from the output URIs; with inlineResponseConfig, read response.results[<audio uri>].transcript.results.
12. Verify against the official docs: https://docs.cloud.google.com/speech-to-text/docs/batch-recognize and https://docs.cloud.google.com/speech-to-text/docs/reference/rest/v2/projects.locations.recognizers/batchRecognize