Travel back in time, even a decade or two ago, and the idea of having a conversation with a smartphone or laptop would be unthinkable to anyone but the earliest of adopters. (For context, Siri didn’t even become a native iPhone feature until 2011.) These days, however, many of us are used to having conversations with Siri, Alexa, Cortana, Bixby, ChatGPT, and others on a daily basis.

Many tech companies are now actively encouraging us to talk to AI assistants in the way that we would to another human. Take Google, for example, which is heavily advertising the ability to hold conversations with Gemini Live everywhere, from billboards and subway posters to TV ads. And speech-to-text (STT) functionality plays a huge role in facilitating that process.

Although AI assistants certainly aren’t the only use for STT — they also have a place in dialog systems, transcription, and assistive tech, among others — the extent to which they rely on them underscores how crucial the technology is to many recent innovations. It’s a functionality that’s increasingly being made available for integration via API.

Below, we’ll compare several major speech-to-text APIs, along with some pros and cons. We’ll explore common use cases for these APIs, touch on how different providers are putting their own spin on STT offerings, and see which might work best for specific needs.

AssemblyAI

With an approach that feels very “API-as-a-product,” — a company profile describes AssemblyAI as “Stripe for AI models” — AssemblyAI’s marketing efforts are actively targeting developers in a big way. This dev-first approach —easy API key generation, extensive documentation, and generous free hours of STT in a playground — will appeal to API creators.

Some of AssemblyAI’s core features, like advanced transcription, auto chaptering and summarization, and content moderation, feel well-suited to podcasts and media indexing. That said, AssemblyAI’s API is being used elsewhere for call transcriptions, to analyze the sentiment of customer inquiries, and much more. They claim to have the industry’s lowest word error rate and up to 30% fewer hallucinations than other providers.

OpenAI Whisper

Given their dominance in AI, it shouldn’t be surprising that OpenAI is looking to extend its influence via Whisper. OpenAI’s Whisper model is open-source and self-hostable for offline applications, and the hosted Speech to text API is an online, paid service. The open-source Whisper model is pre-trained (not fine-tuned for specific domains). Still, it has high accuracy and broad applications.

Whisper was trained on more than 680,000 hours of multilingual data —around ⅓ of Whisper’s dataset is non-English — and was tasked with both transcription and translation of non-English content. With support for more than 100 languages and automatic language detection, it’s a solid choice for developers looking to work with accented or non-English content.

Google Cloud Speech-to-Text

If you already use Google Cloud developer tools, the temptation to use their Speech-to-Text API is probably very real. In this case, it might be worth giving in to that feeling. Google will likely deepen integrations with its suite of products, too, like Gemini, NotebookLM, Text-to-Speech, and Natural Language, over time.

New customers can take it (along with other Google Cloud products) for a spin since it offers free credits to test features like real-time streaming, batch processing, and automatic punctuation for everything from transcription services to voice commands in apps.

Deepgram

Billing themselves as the voice AI platform for enterprise use cases, Deepgram offers APIs for speech-to-text, text-to-speech, voice agents, and audio intelligence. The platform also provides real-time transcription in 36+ languages, custom model training, and topic detection.

Deepgram cites AI agents for use in call centers and conversational AI as their major use cases, with a voice-to-voice API for enabling natural-sounding conversations between humans and machines. Their emphasis on scalability and enterprise use cases, which extends to an Enterprise Accelerator Program, already has Deepgram feeling like the Twilio of voice AI.

Amazon Transcribe

If a dense offering like Deepgram feels like overkill for your project, a transcription offering like Amazon’s speech-to-text service, Transcribe, might be suitable. Though that’s not to say that the product isn’t flexible or suitable for enterprise use — it’s a great option for enterprise software integrated into AWS and for healthcare apps since Amazon Transcribe Medical is covered under AWS’s HIPAA eligibility.

You can also improve accuracy for specific use cases using language customization, content filters (for privacy and/or audience-appropriate language), speech partitioning, and so on. With pay-as-you-go pricing, billed in one-second increments, this is an easy one to experiment with.

IBM Watson Speech to Text

Watson Speech to Text provides pre-trained speech models tuned for the customer service field. Most of the use cases suggested by IBM are in that arena: call analytics, self-service navigation, agent assistance, and speech analytics. The STT service is available as a remote API and containerized library for IBM partners to embed in commercial apps.

With a Lite package that includes 500 minutes of free speech recognition a month and 38 pre-trained speech models, this could be a good option for anyone seeking a low-volume solution, especially if the service you’re looking to build is in the customer care space.

Speechmatics

Founded by Tony Robinson, a researcher in the application of recurrent neural networks to speech, Speechmatics prioritizes accuracy and inclusivity. Their speech-to-text offering is capable of real-time transcription, translation, and summarization, with support for 50+ languages.

Some of their use cases include live captioning, broadcast-grade subtitling, contact center solutions, media monitoring, and education technology. Speechmatics’ focus on improving accessibility and inclusivity might make this an attractive solution for multilingual or diverse projects.

Rev AI

Like Speechmatics, Rev AI emphasizes improving accessibility through diverse datasets. They offer language identification for 20+ languages, along with topic extraction, sentiment analysis, and summarization (for English voice content only).

Rev AI’s site also highlights their comprehensive approach to security — with SOC II, HIPAA, GDPR, and PCI compliance — noting that files are encrypted at rest and in transit. This approach will undoubtedly be attractive to anyone working on projects dealing with sensitive data.

Microsoft Azure Speech-to-Text

Offering APIs like Real-time Speech-to-Text, Fast transcription, and Batch transcription, Microsoft Azure Speech to text feels focused on transcriptions in the same way that Amazon Transcribe does. A few cited use cases include live meeting transcriptions and video subtitling.

If you’re already using related Microsoft products (like Azure App Service or Azure Kubernetes Service) and need an API to add speech-to-text functionality, then Azure STT could offer an easy solution for doing just that.

Vosk API

Alpha Cephei’s Vosk is a speech recognition toolkit that works offline, even on lightweight devices like Raspberry Pi, Android, and iOS. Some notable use cases include IoT devices, embedded systems, wearables, and AR/VR platforms requiring voice input.

Vosk is flexible by nature, and there are plenty of good things to say about it — it’s open-source, privacy-focused, and offers a streaming API — and its low resource usage means it could be a good choice for experimenting with, especially for smaller projects.

Honorable Mentions

Our list of STT APIs only scratches the surface of ongoing development around natural language processing and speech generation. Here are some other noteworthy tools in the realm of speech-to-text and text-to-speech (TTS):

• Twilio Text-to-Speech API: A powerful AI tool within Twilio’s suite of communication APIs.
• Gladia: A comprehensive speech-to-text API, complete with add-ons and vast multilingual support.
• Resemble.ai: Converts text to speech, with audio for real-time conversational agents.
• WellSaid API: Another TTS option that prioritizes human likeness in its voice generation.
• Tavus API: Emphasizes the pairing of TTS and video generation to create lifelike AI avatars.
• Amazon Polly: Another Amazon tool, this time designed to convert text into lifelike speech in dozens of languages.

Takeaways: Choosing The Right One

Whatever your aims, budget, and call volume, there’s an STT API for you. If deep integration is a priority, you might lean towards Google, Microsoft, or Amazon. Vosk or Whisper might be a good way to go for offline projects. Or you could find AssemblyAI’s strong focus on developers an appealing way to test the waters of developing with STT APIs.

But it’s worth pointing out that the speed at which the AI space is evolving is unparalleled. The odds are good that, by the time this article is published, some features or use cases cited as unique selling points will already have been debuted by other providers.

That’s not to say that you shouldn’t spend some time finding an API provider that works for you — it’s merely that STT tools will probably reach a degree of convergence as they look to duplicate each other’s features. And, until then, at least you have plenty of options to choose from.

It’s also interesting to note that, despite their differences, many of these services call out similar use cases: healthcare apps, customer service, and batch transcription and translation. This speaks to the overall direction speech-to-text as a service (STTaaS?) is taking.

It remains to be seen how the rapid evolution of AI will shape that future and how STT will be used beyond those smart agents mentioned in the introduction.

Maybe try, “Hey Gemini, what does the future of speech-to-text look like?”