Speech-to-text (STT) is a crucial element of any voice AI solution. Part of the general Automatic Speech Recognition (ASR) toolbox, STT transcribes spoken words to text, making it possible for voice agents and other Voice AI tools to respond.
Why choosing the right STT provider matters
Whether you’re building real-time voice agents or customer service automation, meeting transcripts, dubbing, captioning, and more, STT is the foundation of nearly every voice-driven experience. That means that picking the right provider for your STT tool is an important part of the process.
How to improve any STT model with real-time speech enhancement
Another significant detail to consider alongside your STT provider is how to set your voice AI tool up to succeed outside of controlled lab conditions. Every STT tool performs better when coupled with AI-powered real-time speech enhancement, which removes background voices, room reverb, and other disruptive real world sounds that break your voice agent.
Some developers turn to de-noising tools like Krisp, but models like Krisp are built to improve audio quality for human ears, not a model’s ears. In fact, by removing subtle phonetic detail as part of the de-noising experience, they often make transcripts less accurate. Ultimately, whichever STT provider you choose, it’s important to couple them with a specific voice enhancement solution, like Quail. Compared to standard de-noising tools Quail delivers consistent improvements in WER (Word Error Rate) .
Specialist vs Cloud STT providers
Specialist STT companies are built around speech as their core product, and they invest heavily in areas that cloud providers don’t optimize for. They also prioritize performance, accuracy, and innovation over broader infrastructure integration. As such, they’re best for real-time AI, high-accuracy, multi-speaker and challenging audio workloads where performance matters more than cloud integration.
In contrast, companies like Amazon, Google, Microsoft and more often STT products as part of their larger cloud ecosystems. Their biggest strengths are integration, scale, security, and enterprise-readiness, but they don’t offer as much fine-tuned flexibility or optimization as the specialist providers. Typically, they’re less useful for Voice AI, so we’ll focus on specialist STT providers today.
Deepgram
Deepgram is a high-performance, developer-friendly STT platform which features both self-hosted and edge options. Key features include:
- Models designed for real-time use
- Streaming and batch support
- Multilingual transcription support available
- Advanced features including speaker diarization and custom vocabulary optimized for voice agents
- Low latency and turn-taking support
- Benchmarks show best performance with Quail
- Quail improves Word Error Rates (WER) by 3-4% – equivalent to a 10-20% drop in total errors
Cartesia
Another tool focused on real-time performance, Cartesia offers real-time APIs especially for conversational voice agents, with features including:
- Extremely low latency
- Curated voices for different conversational needs, or voice cloning on demand
- Multilingual support across 40+ languages
- Integrations available with Vapi, LiveKit and Pipecat
- Benchmarks show best performance with Quail
- Quail reduces incorrect insertions by 1.5-2.5% – equivalent to a 15-25% WER drop
Gladia
Another STT provider optimized for voice agents, Gladia focuses on multilingual access with a rich feature set around diarization and translation such as:
- Support for 100+ languages
- Real-time and asynchronous transcription modes
- Translation, summarisation, and diarization available out of the box
- Low-latency streaming (sub-300ms in ideal conditions)
- Flexible APIs with built-in session metadata, custom prompts and vocabulary
- Benchmarks show best performance with Quail
- Prone to deletions, which Quail corrects by an estimated 20-30%
AssemblyAI
AssemblyAI is focused on high-accuracy STT paired with a comprehensive suite of “speech intelligence” features, including:
- Streaming and batch transcription
- Advanced layers including sentiment analysis, topic detection and PII redaction
- Speaker diarization and summarization options
- Designed for analytics-heavy pipelines in media and call centres
- Benchmarks show best performance with Quail
- Tends to make substitution errors: Quail reduces total errors by an estimated 10-20%
ElevenLabs
Best known for their TTS product, ElevenLabs also provides STT optimised for media, podcasts and content workflows. Features include:
- Scribe v1 model with multilingual support
- Robust multi-speaker diarization (up to 32 speakers)
- Real-time streaming API
- Ideal for long-form or multi-speaker content such as interviews, panels and podcasts
- Benchmarks show best performance with Quail
- Prone to making insertion errors; Quail reduces WER by 15-25%
Try Quail in the ai-coustics SDK
Want to hear how Quail improves your product performance? Try it out today in the ai-coustics SDK – just drop it into your pipeline for immediate transcription accuracy improvements in real-world environments.
Get in touch to speak to an expert and enjoy a personalized demo. Or sign up to our developer platform, obtain your SDK, then clone or download the SDK code from our GitHub repository to start testing it locally.
