Best Speech-to-Text and Text-to-Speech APIs for AI Agents, Independently Tested on 12 Coding Agents (2026)
Lightsage ·
Direct answer
Pick Deepgram to transcribe audio, or AssemblyAI if your team codes with Grok or Pi. Pick ElevenLabs to generate speech, including over MCP. All nine are agent-ready: each scored B or above when 12 coding agents did the same real task in Devtool Arena's October 2026 tests.
A tool is agent-ready here when it scores 75 (B) or higher on Devtool Arena (devtoolarena.com), Lightsage's public leaderboard of how well coding agents use real developer tools. Every tool got the same task, agents and grader (methodology).
The 5 agent-ready speech-to-text APIs
1. Deepgram: 89.3 (B)
Deepgram had the top speech-to-text score.
- Best for: Claude Code, Codex, Cursor and GitHub Copilot
- Best agent result: 97 (A) on Cursor; first on 7 of 12 agents; #1 on both CLI boards (92 and 88)
- Agent assets: all 7 readiness assets, including an MCP server in its CLI (
dg mcp) - Free tier or price: $200 credit, no card (pricing)
2. AssemblyAI: 87.7 (B)
AssemblyAI tells coding agents to use its official assemblyai Python SDK.
- Best for: Grok, Pi, Gemini (Antigravity) and Muse
- Best agent result: 93 (A) on Pi and OpenCode; first on 7 of 12 agents
- Agent assets: all 7 readiness assets, including a hosted MCP server
- Free tier or price: $50 in credits that don't expire; pre-recorded from $0.15 an hour (pricing)
3. OpenAI Whisper: 84.0 (B)
OpenAI's guide recommends gpt-transcribe and says whisper-1 shuts down on February 26, 2027.
- Best for: OpenAI SDK users
- Best agent result: 95 (A) on Hermes; first on 2 of 12 agents
- Agent assets: 6 of 7 readiness assets
- Free tier or price:
gpt-transcribeat an estimated $0.0045 a minute (pricing)
4. Speechmatics: 82.2 (B)
Speechmatics had the strongest average agent run on this page.
- Best for: pre-recorded audio
- Best agent result: 88 (B) on Cursor
- Agent assets: 5 of 7 readiness assets
- Free tier or price: $100 in credits; pre-recorded from $0.12 an hour (pricing)
5. Zoom AI Services: 75.6 (B)
Zoom's Scribe API transcribes in Live, Fast and Batch modes.
- Best for: apps built on Zoom
- Best agent result: 88 (B) on OpenCode
- Agent assets: 6 of 7 readiness assets, including remote MCP servers
- Free tier or price: prepaid credits; Batch at $0.15 an hour (Scribe API)
The 4 agent-ready text-to-speech APIs
1. ElevenLabs: 88.6 (B)
ElevenLabs is the text-to-speech pick for nearly every agent.
- Best for: speech from any coding agent
- Best agent result: 97 (A) on Cursor; first on 11 of 12 agents; #1 on both MCP boards (89)
- Agent assets: all 7 readiness assets, including a hosted MCP server
- Free tier or price: 10K credits a month free; Flash/Turbo at $0.04 per 1K characters (pricing)
2. Rime: 85.4 (B)
Rime's MCP server browses voices and scaffolds voice-agent code.
- Best for: voice agents, and Hermes users
- Best agent result: 92 (A) on Cursor and OpenClaw; first on Hermes (85)
- Agent assets: 6 of 7 readiness assets, including a hosted MCP server
- Free tier or price: $50 of usage a month, no card; Mist v3 at $0.030 a minute (pricing)
3. Speechify: 81.6 (B)
Speechify prices text-to-speech by the character.
- Best for: high-volume speech on a free start
- Best agent result: 90 (A) on Cursor
- Agent assets: 6 of 7 readiness assets, plus a docs MCP server
- Free tier or price: 500K characters a month, no card; Simba 3.5 from $12 per 1M characters (pricing)
4. Cartesia: 79.9 (B)
Cartesia's MCP server can list and clone voices.
- Best for: voice cloning through MCP
- Best agent result: 90 (A) on Cursor; #2 on both MCP boards (80)
- Agent assets: 6 of 7 readiness assets, including a hosted MCP server
- Free tier or price: 20K credits a month free; Pro is $5 a month (pricing)
At a glance
| # | Tool | Category | Score | Best for | Official MCP server |
|---|---|---|---|---|---|
| 1 | Deepgram | Speech-to-text | 89.3 (B) | Most agents | Yes, in its CLI |
| 2 | AssemblyAI | Speech-to-text | 87.7 (B) | Grok and Pi | Yes, hosted |
| 3 | OpenAI Whisper | Speech-to-text | 84.0 (B) | OpenAI SDK | — |
| 4 | Speechmatics | Speech-to-text | 82.2 (B) | Pre-recorded | — |
| 5 | Zoom AI Services | Speech-to-text | 75.6 (B) | Zoom apps | Yes, remote |
| 1 | ElevenLabs | Text-to-speech | 88.6 (B) | Any agent | Yes, hosted |
| 2 | Rime | Text-to-speech | 85.4 (B) | Voice agents | Yes, hosted |
| 3 | Speechify | Text-to-speech | 81.6 (B) | Free volume | Docs only |
| 4 | Cartesia | Text-to-speech | 79.9 (B) | Voice cloning | Yes, hosted |
Prices and MCP details as of November 2026.
How we tested
- Tasks: transcribe a sample MP3 with the vendor's official Python SDK, or save one sentence as output.mp3 using the vendor's docs.
- 12 coding agents, each with a live API key in its own sandbox; same task, agents and grader for every tool.
- Score: 70% agent results, 30% agent-ready assets like llms.txt, an OpenAPI spec and an MCP server; B is 75+. Full method: methodology.
FAQ
What's the best speech-to-text API for AI agents in 2026?
Deepgram, at 89.3 across 12 coding agents in the October 2026 tests, first on 7 of 12 boards.
What's the best text-to-speech API for AI agents in 2026?
ElevenLabs, at 88.6, first on 11 of 12 coding agents' API boards in October 2026.
Is there a voice API that Grok can set up for transcription?
Yes. AssemblyAI ranked first on the Grok board in October 2026 with 90 (A), and Deepgram was second at 86.
Tools can be added at devtoolarena.com/add, and teams that want the same runs on their own tasks can use Lightsage Agent Usability.