What replaced Whisper?

Two things, depending on the job. For batch transcription it is GPT-Transcribe at $0.0045 per minute, which is cheaper than the legacy whisper-1 endpoint and stronger. For live audio it is GPT-Realtime-Whisper at $0.017, which streams natively and removes the chunk-and-stitch layer every live Whisper deployment had to write. Whisper's open weights remain the answer when audio cannot leave your infrastructure.

How much does transcription cost at scale?

About $0.27 per hour of audio on the batch model, so a thousand hours a month is roughly $270. The streaming tier is about $1.02 an hour. The mistake we see most often is a single streaming pipeline handling everything, which turns a $270 line item into a $1,020 one for a latency property only a fraction of the streams actually use.

Which provider is most accurate?

On the Artificial Analysis leaderboard as of June 2026, ElevenLabs Scribe v2 led at 2.2% word error rate against 5.2% for Deepgram Nova-3. That ranking is weaker evidence than it looks: clean English accuracy has plateaued and the differences that matter show up on accents, background noise, overlapping speakers and domain vocabulary. Test on your own recordings.

Do these models do speaker diarisation?

The OpenAI transcription models do not; that is pipeline work, usually pyannote. Google took the other approach: Gemini 3.5 Transcribe carries diarisation, word timestamps and utterance-based language detection in the model at $0.003 per minute of audio. If those are on your requirements list, price both routes before committing to building the layer.

Should we build a voice agent or stitch a pipeline?

Ask whether anyone is waiting. If a human is on the line and can interrupt, use a realtime session: no stitched pipeline matches it on turn latency or overlap handling. If the output is a transcript or a summary, stitch it, because a cheap transcription model plus a strong text model is cheaper and gives you a better text model.

★ Voice & SpeechLLM Models8-week engagement

VOICE · TRANSCRIPTION, AGENTS, SYNTHESIS

Production voice. Transcription, agents, translation, synthesis.

Whisper made transcription a commodity, and the category has moved on twice since. Streaming models replaced chunk-and-stitch, realtime models put reasoning inside the audio loop, and the vendor that leads on word error rate is not the one that leads on turn detection. We build the whole pipeline and route by job.

Realtime APILLM APIEval pipelines

Start a conversation →All llm models →

Cycle

8 weeks · fixed price

Stack

Speech + realtime audio

Output

Production code + eval suite

Handoff

Full source ownership

[THE SHORT VERSION]

Transcription is step one, not the whole job.

Modern speech models transcribe accurately enough that accuracy is rarely what decides a project. The value is in the pipeline: diarisation, timestamps, formatting, redaction, and a clean handoff into summarisation, search, or extraction. The model choice is an afternoon. Building the path that makes the words useful, and proving it on your own recordings rather than on a public leaderboard, is the work.

When it fits

Meeting, call, and media transcription at volume
Voice agents that resolve rather than deflect
Live captions, dictation, and real-time note taking
Cross-language conversation and multilingual support desks
Audio search, quality review, and analysis pipelines

When it does not

Audio that cannot leave your infrastructure, unless you self-host open weights
Anything where you have not settled recording consent for your jurisdictions

[HOW WE BUILD IT]

How we build with Voice & Speech.

Scope and fit

We decide where Voice & Speech earns its place in your system, and where a simpler tool wins. No resume-driven architecture.

Build on a tested foundation

We integrate Voice & Speech against a foundation we trust: typed code, CI, and observability from the first commit. Boring infrastructure, modern surface.

Eval before launch

An eval suite proves the build behaves before it reaches a user. We measure, then ship.

Handoff with ownership

Your team gets the code, the tests, and a runbook. No lock-in to us or to a vendor framework.

[WHAT YOU GET]

What the engagement leaves behind.

Senior

Engineers who have shipped this before

100%

Source ownership at handoff

Eval-first

Tested before it ships

Framework lock-in

[MODELS + VENDORS]

The speech models worth wiring in.

Transcription, speech to speech, translation, and synthesis, across the vendors that lead on each. We integrate them directly behind one abstraction and route by job, because no single provider wins every axis: accuracy, latency, turn detection, language coverage, and cost per hour pull in different directions.

FlagshipOpenAI

May 2026

GPT-Realtime-2.1

gpt-realtime-2.1

Speech to speech, with reasoning and tool use

The voice-agent model. GPT-5-class reasoning with function calling in the loop
128K context, audio and text and image in, audio and text out
Cached input at $0.40 per million, which is the lever that makes it affordable

Read the technical brief

CurrentOpenAI

May 2026

GPT-Realtime-2.1 Mini

gpt-realtime-2.1-mini

Speech to speech, compact tier

Audio in

$10 / 1M

Audio out

$20 / 1M

Roughly a third the audio price of the full 2.1, on the same Realtime API
Text at $0.60 in and $2.40 out, against $4 and $24 on the full tier
Our default for scripted or narrow voice flows that do not need deep reasoning

In the routing table · brief not yet published

FlagshipOpenAI

May 2026

GPT-Transcribe

gpt-transcribe

File transcription, batch

Audio

$0.0045 / minute

The cheapest path to accurate text. About $0.27 per hour of audio
Takes keyword hints and unstructured context for domain terms and code-switching
Serves both the file endpoint and a realtime transcription session

Read the technical brief

CurrentOpenAI

May 2026

GPT-Realtime-Whisper

gpt-realtime-whisper

Streaming speech to text

Audio

$0.017 / minute

Whisper's successor, rebuilt as a streaming model rather than a batch one
Live transcript deltas for captions, notes, and voice interfaces
Just under four times the price of GPT-Transcribe, for latency you can feel

Read the technical brief

CurrentOpenAI

May 2026

GPT-Live-Transcribe

gpt-live-transcribe

Streaming speech to text, tunable

Audio

$0.017 / minute

Same price as Realtime-Whisper, with an explicit latency and accuracy dial
Takes keyword hints, unstructured context, and multiple language hints
Realtime transcription sessions only, no file endpoint

In the routing table · brief not yet published

CurrentOpenAI

May 2026

GPT-Realtime-Translate

gpt-realtime-translate

Speech to speech translation

Audio

$0.034 / minute

Conversational translation that keeps pace with the speaker
More than 70 input languages, 13 output languages
Billed by the minute, so cost per call is predictable before you ship

In the routing table · brief not yet published

CurrentOpenAI

Mar 2025

GPT-4o Mini TTS

gpt-4o-mini-tts

Text to speech

Audio out

$12 / 1M

Text in

$0.60 / 1M

Synthesis for the paths where you generate the words yourself
Cheaper than a full realtime session when there is no conversation to hold
Right for notifications, IVR prompts, and read-aloud

In the routing table · brief not yet published

CurrentGoogle

2026

Gemini 3.5 Transcribe

gemini-3.5-transcribe

Speech to text, with diarisation

Audio in

$0.003 / min

Text out

$0.002 / min

Speaker diarisation and word timestamps in the model, not in your pipeline
Utterance-based language detection for multilingual and code-switching audio
A live variant, gemini-3.5-transcribe-live, for low-latency streams

In the routing table · brief not yet published

CurrentElevenLabs

Nov 2025

ElevenLabs Scribe v2 Realtime

scribe-v2-realtime

Streaming speech to text

Billing

Per minute, tiered

Lowest word error rate on the Artificial Analysis leaderboard as of June 2026, at 2.2%
Sub-150ms latency with more than 90 languages
Where we go when transcript accuracy is the product, not a step in it

In the routing table · brief not yet published

CurrentDeepgram

Apr 2026

Deepgram Flux Multilingual

flux-multilingual

Conversational speech to text

Billing

Per minute, tiered

End-of-turn detection built into the model, so you can drop the external VAD
Median end-of-turn detection under 300ms
The pick when interruption handling decides whether a voice agent feels human

In the routing table · brief not yet published

LegacyOpenAI

Mar 2023

Whisper v1

whisper-1

File transcription, original

Audio

$0.006 / minute

The model that made transcription a commodity, and still self-hostable open weights
Superseded on the API by GPT-Transcribe, which is cheaper and stronger
Keep it only where you need to run the weights on your own hardware

In the routing table · brief not yet published

[METHODOLOGY · K-FRAMEWORK]

Integrated through the
K-Framework.

Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.

Read the K-Framework

Foundations

Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.

Amplification

An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.

Judgment

Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.

[OBSERVABILITY]

Observability your team can read.

A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.

Instrumented

Cost per call

Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.

Instrumented

Latency p50 / p95 / p99

Real distributions, not averages. We know which routes are slow, and why.

Instrumented

Eval pass rates

The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.

Instrumented

Prompt + completion logs

PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.

Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.

[COMMON QUESTIONS]

Questions we get asked.

What replaced Whisper?: Two things, depending on the job. For batch transcription it is GPT-Transcribe at $0.0045 per minute, which is cheaper than the legacy whisper-1 endpoint and stronger. For live audio it is GPT-Realtime-Whisper at $0.017, which streams natively and removes the chunk-and-stitch layer every live Whisper deployment had to write. Whisper's open weights remain the answer when audio cannot leave your infrastructure.
How much does transcription cost at scale?: About $0.27 per hour of audio on the batch model, so a thousand hours a month is roughly $270. The streaming tier is about $1.02 an hour. The mistake we see most often is a single streaming pipeline handling everything, which turns a $270 line item into a $1,020 one for a latency property only a fraction of the streams actually use.
Which provider is most accurate?: On the Artificial Analysis leaderboard as of June 2026, ElevenLabs Scribe v2 led at 2.2% word error rate against 5.2% for Deepgram Nova-3. That ranking is weaker evidence than it looks: clean English accuracy has plateaued and the differences that matter show up on accents, background noise, overlapping speakers and domain vocabulary. Test on your own recordings.
Do these models do speaker diarisation?: The OpenAI transcription models do not; that is pipeline work, usually pyannote. Google took the other approach: Gemini 3.5 Transcribe carries diarisation, word timestamps and utterance-based language detection in the model at $0.003 per minute of audio. If those are on your requirements list, price both routes before committing to building the layer.
Should we build a voice agent or stitch a pipeline?: Ask whether anyone is waiting. If a human is on the line and can interrupt, use a realtime session: no stitched pipeline matches it on turn latency or overlap handling. If the output is a transcript or a summary, stitch it, because a cheap transcription model plus a strong text model is cheaper and gives you a better text model.