Saidwire questions and answers

Saidwire LLC, based in Kansas City, Missouri, makes Wire-1, a real-time English speech-to-text API for voice agents, phone systems and live captions. It has a streaming WebSocket endpoint (wss://api.saidwire.com/v1/listen) and a file API.

Developers who build voice agents and phone agents and need the final transcript quickly after the user stops talking. It plugs into Pipecat, LiveKit Agents and Vapi, and it also serves live captions and file transcription.

35 ms on average and 50 ms at the 95th percentile, 1st of 29 on mean time to final segment (Coval, 30 days to Nov 10, 2026). Coval measures from the end of speech to the final transcript, on its own schedule, against Saidwire's public US-hosted endpoint. Add your own network round trip, and send Finalize when your turn detector fires to get these times.

3.7% word error rate overall and 4.3% on phone-codec audio (Coval, 30 days to Nov 10, 2026). Several vendors are more accurate: the lowest error rates on the same Coval page are 2.2% to 2.5%. Accuracy depends on your audio, so test on your own calls with the 120 free hours. See the comparison.

$0.25 per audio hour at list price, billed per second. Voice platforms and founding customers pay $0.15 per audio hour. You can also prepay: Starter, $25 for 100 audio hours; Growth, $99 a month with 500 audio hours and up to 20 concurrent streams; or a $500 30-day pilot. Pricing page.

Yes. 120 free audio hours, with no card and no deadline, on up to 4 concurrent streams. Get a key at saidwire.com/dev.

For $99 a month: 500 audio hours a month, up to 20 concurrent streams, and priority email support with a reply within one business day. Unused hours do not carry over. Beyond 500 hours, usage is billed at the list rate of $0.25 per audio hour, or contact us for a volume rate.

Payments go through Stripe. After you pay, we add your hours and confirm by email within one business day. Questions: support@saidwire.com.

English only (US, UK and other English accents). For other languages you need a different vendor.

Yes. pip install pipecat-saidwire, then use SaidwireSTTService() as the STT service in your pipeline. Pipecat guide.

Yes. pip install livekit-plugins-saidwire, then pass stt=saidwire.STT() to your AgentSession. LiveKit guide.

Yes, as a custom transcriber. Point the assistant's transcriber at wss://vapi.saidwire.com/vapi/transcriber with your Saidwire key as a Bearer credential. Vapi guide.

Often, yes. Streaming is /v1/listen-compatible and file transcription also works through an OpenAI SDK-compatible endpoint (/v1/audio/transcriptions), so most existing clients switch by changing the base URL and the API key. Parameters Saidwire does not support are accepted and ignored; the list is in the API reference.

Run your own voice-activity or turn detector and send {"type": "Finalize"} the moment the user stops. The final transcript comes straight back. With silence detection alone, a final waits for the pause you set (500 ms or more). The Pipecat and LiveKit plugins send Finalize for you.

Yes. Send 8 kHz μ-law or A-law as it arrives (encoding=mulaw or encoding=alaw), with no transcoding. On phone-codec audio Coval measured 38 ms mean time to final transcript and a 4.3% word error rate (30 days to Nov 10, 2026). Phone guide.

Custom words (keyterm), masking of card, Social Security, phone and email details (redact), numbers written as digits (numerals), server timing on each final (timing) and a filter for faint background speech (noise_guard). Details.

It transcribes English only and one audio channel per stream. It does not add speaker labels or translate. Files are answered in one request, up to 100 MB and 60 minutes each.

No. Audio is processed in memory and never stored, transcripts are not logged, and we do not train on customer data. Security and data.

The API is hosted in the United States. The status page checks it every 5 minutes.

On Coval's public benchmark Wire-1 has the shortest mean time to final transcript (35 ms, 1st of 29; Coval, last measured Nov 10, 2026) and a 3.7% word error rate, which is higher than the most accurate models. Side-by-side figures with sources are on the comparison page.

Support: support@saidwire.com. Founder: Hudson Leach, hudsonleach@saidwire.com. Mail: Saidwire LLC, 215 W Pershing Rd Ste 403 PMB 1093, Kansas City MO 64108.