# Azure audio retirements: whisper, tts and realtime-mini by Dec 15

URL: https://infro.io/blog/azure-audio-retirements-whisper-tts-and-realtime-mini-by-dec-15
Author: INFRO Editorial (https://infro.io/blog/editorial)
Publisher: INFRO
Published: 2026-11-14
Last modified: 2026-11-14
Last reviewed: 2026-11-14 (next review due 2026-12-14)
Topic: Agents and multimodal systems
Format: Guide

> On Azure OpenAI, whisper, tts, tts-hd and gpt-realtime-mini (2025-12-15) retire on 2026-12-15, and the gpt-realtime-2 preview retires on 2027-01-20 with gpt-realtime-2.1 listed as its replacement.[^ms-schedule] Microsoft lists no replacement for the first four, so you pick and test the target yourself, with 31 days left from today.

Key takeaways:
- whisper, tts, tts-hd and gpt-realtime-mini 2025-12-15 retire on 2026-12-15; gpt-realtime-2 retires on 2027-01-20.
- The schedule lists no replacement for whisper, tts, tts-hd or gpt-realtime-mini, so there is no automatic upgrade target and you must choose and test one.
- Retirement dates are not extendable, and a retired model answers 410 Gone, so a missed date is an outage rather than a degradation.
- Voice names, audio formats, turn detection and transcript shape are the places a swap breaks first; latency needs its own measurement per connection method.
- Some candidate versions retire in mid-2027 or have already passed their date, so check the target's own date before you commit.

## What retires, and when

Microsoft's retirement schedule lists these Azure OpenAI audio entries with dates in the next ten weeks.[1]

| Model | Version | Lifecycle | Retirement date | Listed replacement |
| --- | --- | --- | --- | --- |
| whisper | 001 | GA | 2026-12-15 | none |
| tts | 001 | Preview | 2026-12-15 | none |
| tts-hd | 001 | GA | 2026-12-15 | none |
| gpt-realtime-mini | 2025-12-15 | GA | 2026-12-15 | none |
| gpt-realtime-2 | 2026-05-06 | Preview | 2027-01-20 | gpt-realtime-2.1 |

Table: Azure OpenAI audio retirements, from the schedule page as last updated 2026-09-23.

Two reading notes. First, the schedule shows two rows for each gpt-realtime-mini version, one dated in 2026 and one in 2027 (for 2025-12-15: 2026-12-15 and 2027-06-15). The page text does not say what separates them.[1] Until you know which applies, take the earlier date as yours and read the per-SKU `deprecationDate` from the Models API for your own deployments (the call is covered below). Second, the Realtime API page lists gpt-realtime-2 as version 2026-05-07, while the schedule says 2026-05-06.[2] Check the version string on your deployment rather than trusting either page.

From today, November 14, that is 31 days to December 15 and 67 days to January 20.

## What happens on the date

A retired model answers every inference request with 410 Gone, and you can no longer create or keep a deployment of it.[3] Microsoft states that retirement dates are not extendable and that there is no exception process.[4]

Whether anything catches you depends on the row:

- **GA models with a listed replacement.** Global Standard, Data Zone Standard and Standard deployments are auto-upgraded to the replacement on a rolling, region-by-region schedule. `versionUpgradeOption` controls this, and `NoAutoUpgrade` means the deployment stops working at retirement.[4]
- **GA models with no replacement.** whisper, tts-hd and gpt-realtime-mini fall here. Auto-upgrade targets the replacement column, and that column is empty.[1] Expect the deployment to answer 410 Gone.
- **Preview models.** There is no option to stay on a retiring preview. Deployments are force-upgraded to a newer preview or GA version, or terminated if there is no replacement, with at least 30 days of notice.[3] gpt-realtime-2 has a named successor; tts does not.
- **Provisioned deployments.** These are never auto-upgraded, so you migrate them yourself.[3]

Even an auto-upgrade is not a free pass. Microsoft's own guidance is that an endpoint that still responds does not mean the app still behaves correctly, because the traffic starts hitting a different model on a set date.[4] For voice, that difference is audible.

## Pick candidates from what the schedule shows

The schedule names no target for the Dec 15 group, so build a shortlist from entries that are live and have a far date. Read each candidate's own date first; two obvious names are already unsafe.

| Job | Candidate | Date on the schedule | Note |
| --- | --- | --- | --- |
| Speech-to-text | gpt-4o-mini-transcribe 2025-12-15 | 2027-06-15 | Seven months of runway |
| Speech-to-text | gpt-4o-transcribe-diarize 2025-10-15 | 2027-04-15 | Name suggests speaker labels; confirm |
| Speech-to-text | gpt-4o-transcribe 2025-03-20 | 2026-10-15 | Date already passed; do not target |
| Text-to-speech | gpt-4o-mini-tts 2025-12-15 | 2027-06-15 | GA |
| Text-to-speech | gpt-4o-mini-tts 2025-03-20 | 2026-10-15 | Preview, date already passed |
| Realtime | gpt-realtime-2.1-mini | 2027-06-25 | GA |
| Realtime | gpt-realtime-2.1 | 2027-06-25 | GA, also the listed successor of gpt-realtime-2 |
| Realtime | gpt-realtime-1.5 | 2027-08-24 | GA |

Table: Candidate entries and their listed retirement dates.[1]

The Realtime API page also lists gpt-realtime-whisper and gpt-live-transcribe for live transcription, and gpt-realtime-translate for translation, and says they are priced by duration.[2] The schedule table I read has no rows for them, so I cannot give their retirement dates. Look them up in the Models API before choosing one.

The realtime and audio-token models are not drop-in text-to-speech engines, so for the tts and tts-hd jobs decide first whether you want a synthesis model at all. A voice agent already on gpt-realtime-mini may not need one.

Most audio models are only offered as Data Zone or Global deployments, and a successor can appear in some regions before others.[3] Confirm your region and deployment type serve the target before you test, or the test proves nothing.

## Check what your deployments report

Do not rely on the table alone. The Models API returns `lifecycleStatus`, `deprecation.inference` and a per-SKU `deprecationDate` for any model, per subscription and region.[3]

```bash
curl -s -H "Authorization: Bearer $AZ_TOKEN" \
  "https://management.azure.com/subscriptions/$SUB/providers/Microsoft.CognitiveServices/locations/$LOC/models?api-version=2024-10-01" \
  | jq '.value[]
        | select(.model.name | test("whisper|tts|realtime"))
        | {name: .model.name, version: .model.version,
           status: .model.lifecycleStatus,
           retires: .model.deprecation.inference}'
```

Run it once per region you deploy in. Then list every deployment name your code uses, because applications reference the deployment name, not the model. Search code, config and infrastructure templates for those names, including the transcription deployment name that a realtime session passes in `input_audio_transcription`.[2] A realtime session that uses gpt-realtime-mini for conversation and whisper for transcripts has two retirements hiding in one config.

## Test plan: voice, latency and format

Microsoft's migration process says to replay your current configuration on the target unchanged before tuning, so you can separate model-driven change from yours. It also says to freeze the test set for the whole migration.[4] Apply that to audio with three tracks.

### Speech-to-text

Collect recordings from production that you are allowed to keep: phone-band audio, noisy rooms, accents, domain terms and numbers. Freeze them with human reference transcripts. Then compare word error rate and a few targeted checks.

```python
def wer(ref: str, hyp: str) -> float:
    r, h = ref.lower().split(), hyp.lower().split()
    d = [[0] * (len(h) + 1) for _ in range(len(r) + 1)]
    for i in range(len(r) + 1):
        d[i][0] = i
    for j in range(len(h) + 1):
        d[0][j] = j
    for i in range(1, len(r) + 1):
        for j in range(1, len(h) + 1):
            cost = 0 if r[i - 1] == h[j - 1] else 1
            d[i][j] = min(d[i-1][j] + 1, d[i][j-1] + 1, d[i-1][j-1] + cost)
    return d[-1][-1] / max(len(r), 1)

def gate(clips, old, new, max_regression=0.02):
    # clips: [(audio_path, reference_text)]; old/new: callables returning text
    deltas = [wer(ref, new(p)) - wer(ref, old(p)) for p, ref in clips]
    mean = sum(deltas) / len(deltas)
    worst = max(deltas)
    return mean <= max_regression, mean, worst
```

Beyond the average, read the worst clips. Check number and date formatting, proper nouns, language hints (the Realtime API page recommends passing an ISO-639-1 hint for accuracy and latency[2]), and whether your downstream code depends on segment timestamps, word timestamps or a particular response format. If any of those are in use, diff the response shape, not just the text.

### Text-to-speech

Voice is the first thing a listener notices. Voice names are vendor-specific and do not carry over between models or vendors, a point INFRO's audio docs make about fallbacks across speech vendors.[5] So build a listening panel:

- Fix 30 to 50 prompts that include numbers, acronyms, brand names, long sentences and questions.
- Render each on the old and new model with your closest voice, and have reviewers blind-compare.
- Record the output format, sample rate and container for each, and play them through your actual playback path: telephony codecs, browser audio or a mobile player.
- Check clipping at sentence boundaries, speed controls and the behavior on very short input.

If you ship pre-rendered clips (IVR prompts, notifications), regenerate and re-approve them before the date. Cached audio keeps working; new renders will not match it.

### Realtime voice

Here the differences are behavioral. Replay scripted calls on the candidate and compare:

- **Turn handling.** `turn_detection` takes `none`, `semantic_vad` or `server_vad`, and server VAD is the default.[2] Re-tune `threshold`, `prefix_padding_ms` and `silence_duration_ms` on the new model instead of copying values. Watch for early cut-ins and dead air.
- **Session shape.** Use the GA endpoint with `/openai/v1` and no `api-version` parameter.[2] If your client still builds date-versioned URLs, it will fail regardless of model.
- **Tools.** Replay calls that trigger function calls, and compare argument names and ordering.
- **Limits.** The page states 32,000 input tokens and 4,096 output tokens, plus rate limits for audio tokens and concurrent sessions.[2] Check quota on the new deployment, since a new deployment does not inherit your old one.

### Latency, measured per path

The Realtime page quotes about 100 ms for WebRTC and about 200 ms for WebSocket, and SIP as varying.[2] Those are the documented characteristics of the connection method, not a measurement of your calls with the new model. Measure:

- For text-to-speech, time from request to first audio byte, and total render time per 100 characters.
- For transcription, time from end of audio to final text.
- For realtime, time from the server's speech-stopped event to the first audio delta, at p50 and p95, on the same connection method you use in production.

Compare at the tail. A model that is slightly faster at the median and noticeably worse at p95 sounds worse on a phone call.

> **Example: Illustrative acceptance gates**
> These numbers are made up to show the shape of a gate; set your own from your baseline.
> 
> - Mean WER regression at most 2 points absolute on the frozen clip set, and no clip worse than 10 points.
> - p95 realtime response start within 15% of the current deployment.
> - Blind listening panel prefers the new voice or calls it a tie on at least 80% of prompts.
> - Zero format errors in the playback path across the whole prompt set.

## Cut over without a flag day

You have 31 days, so stage it.

1. **This week.** Inventory deployment names and regions, run the Models API check, and confirm the candidate is available in your region and deployment type.
2. **By the end of next week.** Deploy the candidate alongside the old model so rollback stays possible, which is the side-by-side path Microsoft recommends,[4] and run the three test tracks.
3. **The week after.** Send a small share of traffic to the candidate, with the old deployment still live, and watch error rate, p95 latency and any user-facing complaints.
4. **At least five days before December 15.** Complete the move, then delete the old deployment so nothing can still point at it.

For the January 20 preview, the same steps apply with a longer runway. gpt-realtime-2 has a named successor in gpt-realtime-2.1,[1] but preview deployments can be force-upgraded with 30 days of notice,[3] so test the successor now and do not wait for the switch to be made for you.

Past shutdowns show the pattern: Azure's retired list records gpt-4o-realtime-preview 2024-10-01 retiring on March 26, 2025, with a different realtime version named as the replacement.[6] Realtime lines have turned over before, and they will again. The dates in the candidate table above are your next deadline.

## Where a gateway helps, and where it does not

If speech-to-text and text-to-speech calls pass through INFRO, the model lifecycle feature records vendor retirement dates, warns organizations still sending a retiring model, and from the shutdown date refuses the request before any provider call, naming the replacement. That turns a 410 from Azure into an error that says what to switch to. The audio endpoints (`POST /v1/audio/speech` and `POST /v1/audio/transcriptions`) work over Azure OpenAI connections, and a format a target cannot produce is refused before the call so the next target on the route is tried.[5] Failover happens only before the response starts,[7] and since voice IDs do not carry across vendors,[5] keep fallbacks within one voice family. Per-request records show latency and estimated provider cost for each model, which gives you the before-and-after comparison for the latency track. The docs I read describe speech and transcription endpoints; they do not cover realtime sessions, so test those directly against Azure.

## Questions

### Will Azure upgrade my whisper or tts deployment automatically?

Not to anything the schedule names. Automatic upgrades go to the replacement listed in the schedule where one is supported, and whisper, tts, tts-hd and gpt-realtime-mini list none.[1] Treat the deployment as ending on 2026-12-15 and move it yourself.

### Can I get an extension past December 15?

Microsoft's migration guidance says retirement dates are not extendable and that there is no exception process.[4] Plan for the date as published.

## Sources

1. Model retirement schedule — Microsoft Learn. https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule (accessed 2026-11-14)
2. Use the GPT Realtime API for speech and audio — Microsoft Learn. https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/realtime-audio (accessed 2026-11-14)
3. Microsoft Foundry Models lifecycle and support policy — Microsoft Learn. https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirements (accessed 2026-11-14)
4. Model migration: upgrade or switch models in Microsoft Foundry — Microsoft Learn. https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/model-migration (accessed 2026-11-14)
5. Audio & speech — INFRO Docs. https://infro.io/docs/modalities/audio (accessed 2026-11-14)
6. Retired Foundry Models — Microsoft Learn. https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models (accessed 2026-11-14)
7. Failover & fallbacks — INFRO Docs. https://infro.io/docs/features/fallbacks (accessed 2026-11-14)
