The most accurate speech API for developers
Avalon is the batch transcription API tuned for how developers talk. Use the supported OpenAI SDK workflow to get 97.3% accuracy on AISpeak—without replacing your client library.
OpenAI SDK-compatible batch transcription
$0.39 per hour of audio, billed per second with a 10-second minimum
AISPEAK
97.3%
Accuracy
AI jargon, CLI commands, model names. Whisper Large v3 hits 65.1% on AISpeak-10.
OPENASR
6.24%
Average WER
#1 proprietary model on the public OpenASR leaderboard, Oct 2025.
DROP-IN
2-line
Swap
Swap your base URL and model name. Keep your request shape, auth, and tooling.
Why Avalon
People don't speak like an audiobook when they're prompting Claude Code or narrating a deploy. Most training data does. Avalon was trained on real developer workflows, pairing transcripts with CLI sessions and IDE captures, so it keeps the exact command, casing, and model number you said.
Benchmarking alone misses the point. We built Avalon after watching other models fumble obvious technical nouns, turning "zshrc" into "C sharp C" or hallucinating entirely new model names. Avalon stays literal, so your copilots, support analytics, and live demos sound like your users actually talk.
Stop hallucinations on AI terms
AISpeak is our benchmark of real-world clips where people say things like "Claude Code," "GPT-4o-mini," and "o3." Avalon transcribes the key term correctly 97.3% of the time. Whisper Large v3 misses it in more than one out of three attempts.
Accuracy on Coding and AI Terms
Higher is better
AISpeak-10
97.4%
Avalon
Top score
ElevenLabs Scribe
78.8%
Whisper Large v3
65.1%
Voxtral Mini 3B
59.5%
NVIDIA Canary 1B
51.5%
AISpeak-50
97.5%
Avalon
Top score
ElevenLabs Scribe
86.7%
Whisper Large v3
82.4%
Voxtral Mini 3B
79.4%
NVIDIA Canary 1B
71.8%
AISpeak-500
95.8%
Avalon
Top score
ElevenLabs Scribe
87.5%
Whisper Large v3
84.9%
Voxtral Mini 3B
82.9%
NVIDIA Canary 1B
74.1%
Whisper hallucinates. Avalon doesn't.
Avalon
I've
tried
running
this
with
GPT-4o,
GPT-4.1,
and
o3.
Whisper
I've
tried
running
this
with
GPT-4.0,
GPT-4.1,
and
GPT-03.
Avalon
Claude 4
announcement.
Whisper
Claude for
announcement.
Parakeet
cloud floor
announcement.
Avalon
I
just
tried
running
this
with
GPT-4o
and it didn't go very well.
Whisper
I just tried running this with
GPT-400,
and it didn't go very well.
Avalon
Ultrathink.
Whisper
Ultrathing.
Clips pulled from AISpeak evaluations
Proven on industry benchmarks
At its October 2025 debut, Avalon ranked #1 among proprietary models on the public OpenASR leaderboard, #6 overall, at 6.24% average word error rate. Avalon is the same model behind Aqua Voice on Mac, Windows, and iPhone.
Lower is better. Benchmarks measured October 2025 against Whisper Large v3, ElevenLabs Scribe v1, Voxtral Mini 3B, NVIDIA Canary 1B, and AssemblyAI Best.
Lowest average word error rate of any proprietary model on OpenASR at its October 2025 debut.
Beats ElevenLabs Scribe v1 on all six OpenASR datasets shown above.
Keep your OpenAI SDK workflow
Avalon supports the standard OpenAI SDK workflow for its documented batch transcription contract. Change the base URL and model, then check the API reference for exact compatibility.
Python
Node.js
cURL
from openai import OpenAIclient = OpenAI( api_key="your-avalon-api-key", base_url="https://api.aquavoice.com/v1")audio_file = open("speech.mp3", "rb")transcript = client.audio.transcriptions.create( model="avalon-v1.5", file=audio_file,)print(transcript.text)
Works with the official OpenAI SDKs for Python and Node, or plain cURL, for supported batch requests.
Start building with Avalon
Spin up your Avalon API key and launch features where transcripts stay literal, in demos, copilots, and support workflows.
Avalon API
$0.39
/ hour of audio
Everything unlocked. No seat fees.
Billed per second, 10-second minimum
No seat licenses, no enterprise gatekeeping
Files up to 25 MB: mp3, mp4, m4a, wav, webm, mpeg, mpga
Includes segment timestamps and subtitle formats