Compare speech-to-text models on your own audio

One playground for models from many providers

The Speechall playground lists the transcription models in the Speechall catalog. Next to Deepgram (including Nova-3), AssemblyAI and OpenAI (Whisper and GPT Transcribe), the catalog includes models from providers such as ElevenLabs, Speechmatics, Gladia, Microsoft Azure, Rev AI and Groq. Each provider has its own strengths, and the right one depends on your audio. The table lists the models you can pick in the playground, with each model's rate and main features.

A lower rate or a longer feature list doesn't tell you which model transcribes your recordings best. That is what the playground is for.

You don't need an account or API key with each provider. One Speechall account and one prepaid credit balance cover every model in the catalog.

Why test on your own audio

Published benchmarks use someone else's recordings. Your audio has its own accents, vocabulary, microphones, background noise and number of speakers, and those change how each model performs. A model that does well on clean podcast audio can behave differently on a phone call, a noisy meeting room or a lecture full of technical terms.

Pick a few short recordings that represent your real traffic, including the difficult ones. Write a correct transcript for at least one of them, so the word error rate in the playground measures errors and not only the differences between models.

What to compare

• Names, product terms and jargon from your domain, spelled the way you need them
• Numbers, dates, currencies and email addresses
• Punctuation and capitalization, if you show transcripts to people
• Speaker labels on recordings with two or more speakers, including overlapping speech
• Sections with background noise, music, cross-talk or low volume
• Accents and languages your users actually speak, including switches between languages
• Timestamps and SRT or VTT output, if you need subtitles
• Response time and the hourly rate for the volume of audio you expect

WER treats every word the same. Read the diffs as well: one wrong product name can matter more than several missing filler words.

After you choose a model

The playground calls the same Speechall API you integrate with, using the same model IDs, such as deepgram.nova-3, assemblyai.universal-3-5-pro or openai.gpt-transcribe. Put the ID you picked in your request, with the same language, output format and options. To move to a different provider later, change the model ID. Provider-specific options don't all carry over, so check each model's page for the parameters it supports.

• Getting started: create an API key and make your first request.
• Speech-to-text API: audio uploads, remote audio URLs and the OpenAI-compatible endpoint.
• Model catalog: model IDs, languages, features and hourly rates.

Billing works with prepaid credits. You buy credits in the console, and each request, from the playground or the API, draws them down at the model's rate. The billing page shows current purchase fees before checkout.

If you already know you will use one provider and need features only its own API offers, integrating that provider directly can be simpler. Read speech API gateway vs direct integration for the trade-offs.