Switching STT APIs in under 30 minutes

To switch speech-to-text providers safely, put transcription behind an interface, compare outputs on your own audio, and roll out gradually with a fallback. This page is the provider-agnostic playbook. The per-provider guides (Deepgram, AssemblyAI, OpenAI Whisper, ElevenLabs, self-hosted Whisper) cover the exact field mappings.

1. Why teams switch

Accuracy on your audio. General WER doesn't show how a model performs on your domain, accents, and multi-speaker recordings.
Diarization and timestamps. Speaker labels and word times make a transcript searchable and seekable.
Price at volume, output formats (SRT/VTT/DOCX), and operational simplicity (no GPUs to manage).

2. What to evaluate

Dimension	How to judge it
Accuracy	WER on a sample of your own audio, not a vendor's demo clip
Diarization	DER, and whether speaker labels hold up on your own recordings
Timestamps	Word-level start/end accuracy (important for subtitles and search)
Languages	The specific languages and code-switching you serve
Formats and features	JSON shape, SRT/VTT/DOCX, custom vocabulary, webhooks
Throughput and price	Batch latency and cost at your monthly volume

Measure on your data

The public benchmark suite is reproducible and provider-agnostic. Run it on your own audio against the providers you're comparing.

3. Migrate behind an interface

Put transcription behind one function that returns your own normalized shape, instead of calling a vendor SDK throughout your codebase. Switching providers then changes one file. The Speech Revolutions result exposes .text, .words, and .utterances, and to_deepgram() returns a Deepgram-shaped dict if you are migrating from Deepgram.

# transcription.py — the ONE place your app calls
from speechrevolutions import SpeechRevolutions

_client = SpeechRevolutions()  # SPEECHREVOLUTIONS_API_KEY

def transcribe(audio: str) -> dict:
    r = _client.transcribe(audio, speaker_labels=True, word_timestamps=True)
    return {
        "text": r.text,
        "speakers": [{"speaker": u.speaker, "text": u.text} for u in r.utterances],
        "words": [w.to_dict() for w in r.words],
    }

4. Compare outputs before you cut over

Run both providers on the same sample set through your interface, store both outputs, and compare them. Measure WER against a reference transcript if you have one; otherwise, spot-check the transcripts, speaker turns, and timestamps your product depends on. Keep the sample as a regression set.

5. Roll out incrementally

Shadow. Send a copy of production traffic to Speech Revolutions and compare results without serving them.
Ramp. Route 5% → 25% → 100% of traffic behind a feature flag, and check quality and error metrics at each step.
Fall back. On error or timeout, have the interface call the old provider until the rollout is complete.

For large backfills

To re-transcribe an existing library, use submit() with polling or webhooks instead of blocking calls. See the batch tutorial.

6. The 30-minute quickstart

Install the SDK, set your API key, transcribe, and read the result.

pip install speechrevolutions
export SPEECHREVOLUTIONS_API_KEY=stt_...

from speechrevolutions import SpeechRevolutions

client = SpeechRevolutions()
result = client.transcribe("meeting.mp3", speaker_labels=True)
print(result.text)

Next, see the provider-specific guide for exact field mappings, or the Quickstart.