Compare

Speech-to-text APIs for voice agents compared

On Coval's public speech-to-text benchmark, Wire-1 by Saidwire has the shortest mean time to final transcript of the 29 models ranked, 35 ms (Coval, last measured Nov 10, 2026), but it is not the most accurate: its word error rate is 3.7%, and the lowest on Coval's page are 2.2% to 2.5% (Coval word error rate page, read Nov 10, 2026). No API leads on speed, accuracy, languages and price at once, so the right one depends on what your voice agent needs most.

Side by side

Five vendors a voice-agent builder often weighs, with every model Coval lists for each. Lower is better in both number columns. Coval ranks more vendors than this page covers; its pages have the full list.

Speed and error rate: Coval's time to final segment and word error rate pages, rolling 30 days, read Nov 10, 2026. Saidwire's row covers 3,616 samples. Prices: each vendor's own pricing page, read Nov 10, 2026.

How to choose

• Shortest wait after the user stops talking: Wire-1 (35 ms average, 50 ms p95; Coval, last measured Nov 10, 2026). English only.
• Lowest error rate: not Saidwire. The most accurate models on Coval's page measure 2.2% to 2.5%, against 3.7% for Wire-1 (Coval word error rate page, read Nov 10, 2026). Of the vendors on this page, Soniox STT RT v6 is slightly ahead of Wire-1 at 3.6% (Coval, last measured Nov 10, 2026). If a wrong word costs you more than a pause, start with Coval's accuracy ranking.
• Languages other than English: not Saidwire. Wire-1 is English only; most of the other vendors here list more languages on their own pricing pages.
• Speaker labels, translation or text-to-speech from the same vendor: not Saidwire, which sells speech-to-text only. See each vendor's page.
• Price: compare the listed prices above with your own volume. Saidwire's list price is $0.25 per audio hour, $0.15 for voice platforms and founding customers, with 120 free hours and no card. Three of the four other vendors on this page list a lower hourly price than that; see the table.

One-to-one comparisons

• Saidwire vs Soniox
• Saidwire vs ElevenLabs
• Saidwire vs Speechmatics
• Saidwire vs OpenAI

How to read these numbers

• Who measured. Speed and error-rate figures are Coval's, from its public benchmark pages. Coval runs every model daily on the same audio and starts every timer from one shared end-of-speech point. Saidwire ran none of these tests and does not test other vendors.
• Time to final segment is the wait between the end of speech and the final transcript: the wait a voice agent has before it can answer. It leaves out your network round trip and each vendor's own turn-detection delay.
• Word error rate is the share of words a transcript gets wrong on Coval's test audio. Your audio will differ.
• Prices are list prices read from each vendor's public pricing page on Nov 10, 2026. Plans, discounts and add-ons change them; the vendor's page is the authority.
• What Coval leaves out. Coval does not score every model, language, add-on or setting a vendor offers, and it says so on each provider page. Features such as speaker labels, translation and language coverage are not part of these two numbers.
• Figures roll daily. Each row links to the page it came from, with that page's date.

Test on your own audio

Benchmarks narrow the list; your own calls decide. Saidwire's free key covers 120 audio hours with no card (saidwire.com/dev). Stream a call and time the gap between Finalize and the final result:

# pip install websockets   (Python 3.10+)
import asyncio, json, os
from websockets.asyncio.client import connect

URL = ("wss://api.saidwire.com/v1/listen?encoding=linear16&sample_rate=16000"
       "&interim_results=true&punctuate=true&endpointing=false")

async def run(frames):                     # frames: 20-100 ms chunks of 16-bit PCM
    headers = {"Authorization": f"Token {os.environ['SAIDWIRE_API_KEY']}"}
    async with connect(URL, additional_headers=headers) as ws:
        async def read():
            async for raw in ws:
                m = json.loads(raw)
                if m.get("type") == "Results":
                    text = m["channel"]["alternatives"][0]["transcript"]
                    print("final  " if m["is_final"] else "interim", text)
        reader = asyncio.create_task(read())
        async for chunk in frames:
            await ws.send(chunk)
        # your turn detector says the user stopped: ask for the final now
        await ws.send(json.dumps({"type": "Finalize"}))
        await ws.send(json.dumps({"type": "CloseStream"}))
        await reader

Run the same audio through the other vendors you are weighing with their own keys and SDKs, under their terms, and compare what you see.

Questions

Questions or stuck? Email support@saidwire.com. Saidwire LLC, 215 W Pershing Rd Ste 403 PMB 1093, Kansas City MO 64108.

Soniox, ElevenLabs, Speechmatics, OpenAI and the other names on this page belong to their respective owners. They are named only to describe compatibility or to compare public figures; no partnership or endorsement is implied. Coval benchmark figures are from Coval's public pages; Coval does not endorse Saidwire.