What is IBM Watson Speech to Text?

IBM Watson Speech to Text is IBM's cloud speech recognition service. You send it audio through an API, and it returns a transcript. Its free Lite plan includes 500 minutes of recognition a month, with 38 pre-trained speech models ¹. In one 2024 academic test on English university lecture recordings, IBM had an average word error rate of 11.2 percent, about 3.9 times the rate of Whisper large-v2 ^{2 3}. That is one test from 2024, so check the current service on your own audio. For the other providers, see the guide to speech-to-text APIs.

IBM Watson Speech to Text is a cloud API, not an app

The service runs in IBM Cloud. IBM's product page describes it as a way to convert speech into text with AI-powered recognition and transcription ( IBM Speech to Text product page). Developers call it from their own software, and you do not install it on a computer. You create a service instance, send audio to its recognition endpoint, and read the JSON that comes back. IBM documents the recognition method in its online API reference, with the list of models and the request options ( IBM Cloud API reference).

This is the same kind of service as Azure Speech to text or Google Speech-to-Text. The difference is in the plans, the models and the accuracy you measure.

Watson is IBM's name for several AI services. The speech to text service is the one that transcribes audio, and it is separate from the text to speech service, which reads text aloud.

Plans: a free Lite tier, then paid plans with more capacity

IBM sells the service in tiers. The Lite plan is free and includes 500 minutes of recognition a month, with 38 pre-trained models ¹. The Plus plan includes unlimited minutes a month and 100 concurrent transcriptions ⁴. IBM's product page also lists a Premium plan for large and security-sensitive firms, with more capacity and data protection ( IBM Speech to Text product page).

We could not read the per-minute rates for the paid plans in the text of IBM's product or pricing pages, so this page does not give an IBM price per minute. Check the plan you want in the IBM Cloud catalog before you budget, and calculate the minutes you expect to send each month. The guide to speech-to-text APIs lists the per-hour prices of the providers we could verify.

Accuracy: IBM's 2024 result was behind Whisper on lecture audio

The clearest independent number we found is from a 2024 study that measured the accuracy of several automatic speech recognition services on university lecture recordings. In its English results, IBM had an average word error rate of 11.2 percent ². In the same study, Whisper large-v2 had an average of 2.9 percent ⁵. That puts IBM at about 3.9 times Whisper's error rate on that set ³.

Read that result with its limits in mind. It covers one kind of audio, lectures, tested in 2024. IBM may have changed its models since then, and a lecture is not a phone call or a meeting with several speakers. The guide to speech recognition accuracy explains why results from one test do not carry over to every recording.

It depends on whether you need a cloud API or a model you run yourself

• Choose IBM Watson when you already use IBM Cloud, when you want to tune a model with your own vocabulary, or when your procurement team needs IBM's contract and data terms. IBM's product page says its Plus plan lets you tune speech models to improve accuracy in recognition ( IBM Speech to Text product page).
• Choose another API when accuracy on your audio is the main test, since the 2024 result favored Whisper on lectures. A test on your own recordings will settle it faster than a brochure.
• Choose a free local model when audio must stay on your machine or the volume is high. The question on whether Whisper is free covers the costs of running it yourself.

What to do

1. Create a free Lite instance and send it a five-minute recording with your names and terms in it.
2. Compare the transcript with a hand-checked passage, and count the word errors. The word error rate guide explains the calculation.
3. Run the same audio through one other API and one local model, using the same settings.
4. Price the plan you need from the IBM Cloud catalog, using your expected monthly minutes.

Related answers: how to use the Azure Speech-to-Text API, how to use the ElevenLabs speech-to-text API, whether there is a free speech-to-text API and whether ElevenLabs is better than Whisper. The overview of speech recognition explains the technology.