Free transcription models today
The table lists the free transcription models in the FreeLLMAPI catalog. The data comes from the live catalog. "Max upload" is the largest file that the provider accepts on the free tier, when the catalog knows it.
| Model | Provider | Max upload | Free limit |
|---|---|---|---|
Whisper Large v3 Turbo (Groq)whisper-large-v3-turbo | Groq | 25 MB | free tier - 25 MB uploads |
Whisper Large v3 (Groq)whisper-large-v3 | Groq | 25 MB | free tier - 25 MB uploads |
Whisper Large v3 Turbo (Cloudflare Workers AI)@cf/openai/whisper-large-v3-turbo | Cloudflare Workers AI | - | Workers AI free allocation |
Whisper (Cloudflare Workers AI)@cf/openai/whisper | Cloudflare Workers AI | - | Workers AI free allocation |
Deepgram Nova-3 (Cloudflare Workers AI)@cf/deepgram/nova-3 | Cloudflare Workers AI | - | Workers AI free allocation |
Whisper Tiny English (Cloudflare Workers AI)@cf/openai/whisper-tiny-en | Cloudflare Workers AI | - | Workers AI free allocation |
Cohere Transcribe Arabic (07-2026)cohere-transcribe-arabic-07-2026 | Cohere | 25 MB | free evaluation key - 1,000 calls/month shared |
Cohere Transcribe (03-2026)cohere-transcribe-03-2026 | Cohere | 25 MB | Free evaluation key — 1,000 calls/month shared; language defaults to English |
Voxtral Mini Transcribe 2 (Mistral)voxtral-mini-2602 | Mistral | 25 MB | Renewing Mistral Free-mode monthly usage shared with Studio/API/Vibe; account-specific allowance. Paid overage may apply if enabled. |
Whisper Large v3 (OVH)whisper-large-v3 | OVH AI Endpoints | 25 MB | free · anonymous 2 req/min per IP |
Whisper Large v3 Turbo (OVH)whisper-large-v3-turbo | OVH AI Endpoints | 25 MB | free · anonymous 2 req/min per IP |
Typhoon ASR Realtimetyphoon-asr-realtime | Typhoon | - | Ongoing free research API (shared rate limits) |
Typhoon Isan ASR Realtimetyphoon-isan-asr-realtime | Typhoon | - | Ongoing free research API (shared rate limits) |
Live data from the FreeLLMAPI catalog (version 2026.11.19, checked 19 November 2026). Free limits change often. The catalog is checked every day.
Most entries are Whisper models from OpenAI, hosted by different providers. The table also has Deepgram Nova-3 on Cloudflare, Voxtral from Mistral, Cohere Transcribe, and Typhoon ASR for Thai. For a comparison of the Whisper models only, read Free Whisper API.
Use this rule to choose a model:
- General transcription in many languages: Whisper Large v3 or Whisper Large v3 Turbo. Turbo is faster. Large v3 has a lower word error rate on Groq.
- Subtitles: a Cloudflare Whisper model, because it returns VTT.
- Arabic: Cohere Transcribe Arabic. Thai: Typhoon ASR.
- Fallback: OVH Whisper. It needs no key, but it allows only 2 requests per minute.
Step 1: Install FreeLLMAPI and add keys
FreeLLMAPI is open source. It runs on your computer in Docker. Your provider keys stay on your computer.
Install (macOS or Linux)
curl -fsSL https://freellmapi.co/install.sh | bashOpen the dashboard at http://localhost:3001. Add a Groq key and a Cloudflare key on the Keys page. Both providers give free keys, and both serve Whisper. Copy your unified key from the same page.
account_id:token. A token without the account ID does not work.Step 2: Send an audio file
The endpoint takes a multipart/form-data upload, the same as the OpenAI API. Two fields are required: file and model. Use model=auto or model=whisper-1 to let the router select a provider.
curl
curl http://localhost:3001/v1/audio/transcriptions \
-H "Authorization: Bearer <unified-key>" \
-F [email protected] \
-F model=auto \
-F language=en \
-F response_format=jsonPython (OpenAI SDK)
from openai import OpenAI
client = OpenAI(base_url="http://localhost:3001/v1", api_key="<unified-key>")
with open("meeting.mp3", "rb") as audio:
result = client.audio.transcriptions.create(
model="auto", # or a model id, e.g. "whisper-large-v3-turbo"
file=audio,
language="en", # optional ISO-639-1 code; improves accuracy
prompt="FreeLLMAPI, Groq", # optional: names and terms in the audio
)
print(result.text)The response has an X-Provider header and an X-Model header. These headers show which provider and model made the transcript.
Options and output formats
| Field | Values | Notes |
|---|---|---|
file | One audio file | Required. Maximum 25 MB for each upload through FreeLLMAPI |
model | auto, whisper-1 or a model id | Required. A model id pins one provider row |
response_format | json, text, verbose_json, vtt | Default is json. srt returns error 400 |
language | Language code, for example en | Optional. Set it when you know the language |
prompt | Text | Optional. Give names and special terms |
temperature | 0 to 1 | Optional. Keep 0 for stable output |
- verbose_json gives segments with times when the provider returns segments. If the provider has no segments, you get the plain JSON shape.
- vtt subtitles come only from providers that make them natively. In the catalog, these are the Cloudflare Whisper models.
- srt is refused. No configured provider makes SRT natively, and the gateway does not convert it.
Free limits and how to stay inside them
Free transcription limits use different units. Some providers count requests. Others count audio time or a daily compute budget.
| Provider | Free limit (official) | What it means for you |
|---|---|---|
| Groq | 25 MB for each file on the free tier. Requests under 10 seconds are billed as 10 seconds | Fast. Good first choice for short and medium files |
| Cloudflare Workers AI | 10,000 neurons each day. Whisper Large v3 Turbo uses 46.63 neurons for each audio minute | About 214 audio minutes each day, if no other Workers AI model uses the budget |
| OVH AI Endpoints | 2 requests per minute for each IP address and model, without a key | Slow fallback. Do not use it for batches |
| Cohere | Trial key: 1,000 API calls each month for all endpoints | Calls are shared with chat and embeddings |
- Add keys from two or more providers. When Groq sends a 429 error, the router sends the file to the next provider.
- Convert large files before upload. Groq says that its models downsample audio to 16 kHz mono. Do this yourself to make the file smaller.
- Split audio longer than about 25 MB into parts. Transcribe each part, then join the text.
Make a file smaller with ffmpeg (16 kHz mono FLAC)
ffmpeg -i interview.m4a -ar 16000 -ac 1 -c:a flac interview.flacTroubleshooting
| Problem | Cause | Fix |
|---|---|---|
Error 413 file_too_large | The file is larger than 25 MB | Convert to 16 kHz mono, or split the file |
Error 400 model is required | The form has no model field | Add -F model=auto |
Error 400 unsupported_format | You asked for srt | Use vtt, json, text or verbose_json |
Error 503 no_transcription_models | No transcription model is enabled | Add a Groq or Cloudflare key, then wait for the catalog sync |
| Wrong words for names | The model does not know the terms | Send the names in the prompt field |
Empty vtt output | The provider does not make VTT | Pin a Cloudflare Whisper model id |
Frequently asked questions
Is there a free speech to text API?
Yes. Groq, Cloudflare Workers AI, OVH, Mistral and Cohere give free transcription with limits. FreeLLMAPI puts 13 free models behind one OpenAI-compatible endpoint.
What is the maximum audio file size?
FreeLLMAPI accepts uploads of up to 25 MB. Groq also allows 25 MB on its free tier. For longer audio, convert it to 16 kHz mono or split it into parts.
Can I get subtitles from a free speech to text API?
Yes. Set response_format to vtt. Only the Cloudflare Whisper models make VTT natively. SRT is not available.
Does the free speech to text API work with the OpenAI Python SDK?
Yes. Set base_url to http://localhost:3001/v1 and call client.audio.transcriptions.create with a file and model="auto".
Related guides
Sources
- FreeLLMAPI README: supported endpoints (checked 20 November 2026)
- Groq docs: Speech to text (checked 20 November 2026)
- Cloudflare Workers AI pricing (checked 20 November 2026)
- Cloudflare Workers AI: whisper-large-v3-turbo (checked 20 November 2026)
- OVHcloud AI Endpoints capabilities (checked 20 November 2026)
- Cohere docs: rate limits (checked 20 November 2026)