Comparison
Best Whisper APIs for Long Audio (2026): 9 Options Compared
MyAi Network · · 9 min read
Most speech-to-text comparisons rank APIs on a 30-second clip. Long audio is a different job: hour-long podcast episodes, lecture recordings, a back catalog of hundreds of files. What matters then is whether the API takes a 2-hour file at all, whether it can fetch audio from a URL, whether it calls you back when it is done, and whether it gives you speaker labels and caption files without extra work.
We run one of these services, MyAi, so read this with that in mind. We have tried to be fair: three of the vendors below list a lower price per hour than we do, and the article says so. Every figure comes from the vendor's own pricing or documentation page, opened on October 14, 2026, and is linked. Where we could not confirm something on the vendor's own page, we say so instead of guessing. Not every option runs Whisper: AssemblyAI, Deepgram's Nova-3 and OpenAI's gpt-4o models are their own models, included because people comparing Whisper APIs compare them too.
The comparison table
Prices are for batch or pre-recorded audio, not live streaming, converted to dollars per audio hour. Checked October 14, 2026. Prices change; check the source before you commit.
| Service | Price per audio hour | File size / length | URL and batch input | Webhooks | Speaker labels | SRT / VTT |
|---|---|---|---|---|---|---|
| Groq | $0.04/hr (turbo), $0.111/hr (large-v3); Batch API 50% off (source) | 25 MB per file (free tier), 100 MB (dev tier); 10-second minimum billed per request | One file per request, upload or a url parameter. Batch API accepts transcription jobs (24-hour to 7-day window, 50% discount) | Not documented | Not documented | No SRT/VTT; json, verbose_json (segment and word timestamps) or text |
| Salad Transcription API | $0.20/hr (0–8,333 h/mo), $0.18, $0.16 at volume; Lite $0.08 / $0.05 / $0.03 (source) | Billed per 0.01 hour (36-second minimum). Jobs are kept under 2 hours of processing time; maximum file size not stated on the pages we read | Media URL (required input field); one job per file | Yes (optional webhook per job) | Yes | SRT (set srt: true); VTT not documented |
| Together AI | $0.0015/min = $0.09/hr (Whisper Large v3) (source) | 80 MB direct upload, 1 GB by URL, 4 hours of audio per request | Upload, or a public HTTPS URL as the file field (the Python SDK wants a local file) | Not documented | Yes (diarize=true, SPEAKER_00 style labels); no extra charge mentioned | json and verbose_json (word and segment timestamps); SRT/VTT not documented |
| MyAi (us) | $0.10/hr, billed to the second of accepted audio; 10 free hours (source) | Files up to 4 hours each through the API (split into 10-minute windows and stitched back together); uploads up to 2 GB on the no-code page | One call with a list of audio URLs; the no-code page also takes a podcast RSS feed or episode page | Yes (webhook_url on the batch), or poll | Yes ("diarize": true), included; name speakers once and every format uses the names | SRT, VTT, TXT and JSON with sentence and word timestamps, included |
| AssemblyAI | $0.15/hr (Universal-2), $0.21/hr (Universal-3.5 Pro); speaker labels +$0.02/hr (source) | Up to 5 GB per request by URL (2.2 GB via their upload endpoint); 160 ms to 10 hours per file | Public audio_url or upload; one transcript per request | Yes (webhook_url; retried up to 10 attempts) | Yes (speaker_labels), +$0.02/hr standard or +$0.065/hr experimental | SRT and VTT export endpoints |
| OpenAI | $0.003/min = $0.18/hr (mini); $0.006/min = $0.36/hr (gpt-4o-transcribe, -diarize, whisper-1) (source) | 25 MB per file; split longer recordings into chunks yourself | File upload only; the Batch API does not list the transcription endpoint | No transcription webhook documented | Only with gpt-4o-transcribe-diarize (diarized_json) | SRT and VTT on whisper-1 only; gpt-4o models return json or text |
| Deepgram | Nova-3 $0.22/hr (Growth) to $0.26/hr (pay as you go); multilingual $0.26–$0.31; Whisper Large $0.29/hr (source) | 2 GB per file; requests that take over 10 minutes to process (20 minutes for Whisper) time out | Hosted URL or upload; one file per request | Yes (callback URL) | Yes, included for pre-recorded audio | SRT and WebVTT through Deepgram's caption libraries, converted from the JSON response |
| Replicate (openai/whisper) | No per-audio-hour price: billed for GPU time (T4 at $0.000225/sec); the model page estimates about $0.0012 per run (source) | Local uploads through the client up to 100 MB; URL size limit not documented | Hosted URL or file, one prediction per file | Yes (prediction webhooks) | Not offered by this model | Output as plain text, SRT or VTT |
| Fireworks AI | Could not verify: no speech-to-text price on the pricing page (source) | Could not verify | Could not verify | Could not verify | Could not verify | Could not verify |
Want the totals for your own volume? The transcription cost calculator multiplies these list prices by your hours of audio.
The list, lowest listed price first
1. Groq
- Price: $0.04/hr (turbo), $0.111/hr (large-v3); Batch API 50% off. Source: console.groq.com/docs/models.
- Models: whisper-large-v3-turbo, whisper-large-v3.
- Long files: 25 MB per file (free tier), 100 MB (dev tier); 10-second minimum billed per request.
- Input: One file per request, upload or a url parameter. Batch API accepts transcription jobs (24-hour to 7-day window, 50% discount).
- Webhooks: Not documented. Speaker labels: Not documented. Captions: No SRT/VTT; json, verbose_json (segment and word timestamps) or text.
- Signup credit: Not counted (free tier is rate-limited, not a credit).
- Best for: Lowest price per hour when your files fit under the size cap or you will compress and cut them yourself.
The same Whisper large-v3-turbo model MyAi runs, for less per hour. Long recordings have to be compressed or split to fit 25 MB or 100 MB, and you stitch the pieces back together. Docs: Groq speech-to-text docs.
2. Salad Transcription API
- Price: $0.20/hr (0–8,333 h/mo), $0.18, $0.16 at volume; Lite $0.08 / $0.05 / $0.03. Source: docs.salad.com billing.
- Models: Salad Transcription API and Transcription API Lite.
- Long files: Billed per 0.01 hour (36-second minimum). Jobs are kept under 2 hours of processing time; maximum file size not stated on the pages we read.
- Input: Media URL (required input field); one job per file.
- Webhooks: Yes (optional webhook per job). Speaker labels: Yes. Captions: SRT (set srt: true); VTT not documented.
- Signup credit: Not stated on the pages we read.
- Best for: Very large monthly volumes where the tiered rate drops, or the Lite tier when you only need plain transcripts.
Runs on a distributed GPU network, as MyAi does. Lite is cheaper per hour than MyAi; check which features Lite drops before choosing it. Docs: Salad Transcription overview. Not verified on October 14, 2026: maximum file size, signup credit.
3. Together AI
- Price: $0.0015/min = $0.09/hr (Whisper Large v3). Source: together.ai/pricing.
- Models: Whisper Large v3 (also NVIDIA Parakeet and Nemotron ASR).
- Long files: 80 MB direct upload, 1 GB by URL, 4 hours of audio per request.
- Input: Upload, or a public HTTPS URL as the file field (the Python SDK wants a local file).
- Webhooks: Not documented. Speaker labels: Yes (diarize=true, SPEAKER_00 style labels); no extra charge mentioned. Captions: json and verbose_json (word and segment timestamps); SRT/VTT not documented.
- Signup credit: Not checked.
- Best for: Per-request Whisper with larger files than Groq accepts, if you are happy to build captions yourself.
Slightly lower per hour than MyAi. One request per file; you manage concurrency and retries. Docs: Together speech-to-text docs. Not verified on October 14, 2026: signup credit.
4. MyAi (us)
- Price: $0.10/hr, billed to the second of accepted audio; 10 free hours. Source: myaitoken.io/transcribe.
- Models: whisper-large-v3-turbo.
- Long files: Files up to 4 hours each through the API (split into 10-minute windows and stitched back together); uploads up to 2 GB on the no-code page.
- Input: One call with a list of audio URLs; the no-code page also takes a podcast RSS feed or episode page.
- Webhooks: Yes (webhook_url on the batch), or poll. Speaker labels: Yes ("diarize": true), included; name speakers once and every format uses the names. Captions: SRT, VTT, TXT and JSON with sentence and word timestamps, included.
- Signup credit: 10 audio hours per email address.
- Best for: Back catalogs of long public recordings (podcasts, lectures, archives) sent as a list of URLs, with captions and speaker labels in one place.
Not the lowest price per hour. Audio is processed on GPUs run by independent operators, so do not send confidential, medical, legal or regulated recordings. Docs: MyAi transcription API.
5. AssemblyAI
- Price: $0.15/hr (Universal-2), $0.21/hr (Universal-3.5 Pro); speaker labels +$0.02/hr. Source: assemblyai.com/pricing.
- Models: Universal-2, Universal-3.5 Pro.
- Long files: Up to 5 GB per request by URL (2.2 GB via their upload endpoint); 160 ms to 10 hours per file.
- Input: Public audio_url or upload; one transcript per request.
- Webhooks: Yes (webhook_url; retried up to 10 attempts). Speaker labels: Yes (speaker_labels), +$0.02/hr standard or +$0.065/hr experimental. Captions: SRT and VTT export endpoints.
- Signup credit: $50, no card.
- Best for: The longest single files (up to 10 hours) with captions, webhooks and extra speech features.
Not Whisper: AssemblyAI runs its own models. Strong all-round long-audio option. Docs: AssemblyAI transcription docs.
6. OpenAI
- Price: $0.003/min = $0.18/hr (mini); $0.006/min = $0.36/hr (gpt-4o-transcribe, -diarize, whisper-1). Source: OpenAI API pricing.
- Models: gpt-4o-mini-transcribe, gpt-4o-transcribe, gpt-4o-transcribe-diarize, whisper-1.
- Long files: 25 MB per file; split longer recordings into chunks yourself.
- Input: File upload only; the Batch API does not list the transcription endpoint.
- Webhooks: No transcription webhook documented. Speaker labels: Only with gpt-4o-transcribe-diarize (diarized_json). Captions: SRT and VTT on whisper-1 only; gpt-4o models return json or text.
- Signup credit: Not checked.
- Best for: Short files where you already use the OpenAI API.
The 25 MB cap means an hour-long WAV must be compressed or cut first. Docs: OpenAI speech-to-text guide. Not verified on October 14, 2026: signup credit.
7. Deepgram
- Price: Nova-3 $0.22/hr (Growth) to $0.26/hr (pay as you go); multilingual $0.26–$0.31; Whisper Large $0.29/hr. Source: deepgram.com/pricing.
- Models: Nova-3, Whisper Large.
- Long files: 2 GB per file; requests that take over 10 minutes to process (20 minutes for Whisper) time out.
- Input: Hosted URL or upload; one file per request.
- Webhooks: Yes (callback URL). Speaker labels: Yes, included for pre-recorded audio. Captions: SRT and WebVTT through Deepgram's caption libraries, converted from the JSON response.
- Signup credit: $200.
- Best for: Large files with diarization included, and the biggest signup credit on this list.
Not Whisper by default (Nova-3); Whisper Large is offered at a higher rate. The $200 credit covers more hours than our 10 free hours. Docs: Deepgram pre-recorded docs.
8. Replicate (openai/whisper)
- Price: No per-audio-hour price: billed for GPU time (T4 at $0.000225/sec); the model page estimates about $0.0012 per run. Source: replicate.com/pricing.
- Models: openai/whisper (community-hosted).
- Long files: Local uploads through the client up to 100 MB; URL size limit not documented.
- Input: Hosted URL or file, one prediction per file.
- Webhooks: Yes (prediction webhooks). Speaker labels: Not offered by this model. Captions: Output as plain text, SRT or VTT.
- Signup credit: Not checked.
- Best for: Experiments and prototypes where you want to swap between Whisper variants.
Cost depends on how long the GPU runs, so it cannot be converted to a per-hour list price honestly. Measure it on your own files. Docs: replicate.com/openai/whisper. Not verified on October 14, 2026: cost per audio hour, maximum file size by url.
9. Fireworks AI
- Price: Could not verify: no speech-to-text price on the pricing page. Source: fireworks.ai/pricing.
- Models: Previously listed Whisper V3 and V3 Turbo.
- Long files: Could not verify.
- Input: Could not verify.
- Webhooks: Could not verify. Speaker labels: Could not verify. Captions: Could not verify.
- Signup credit: Could not verify.
- Best for: Check with Fireworks directly.
On October 14, 2026 Fireworks' pricing page listed no audio pricing and its Whisper model pages returned 404. We have left it out of the cost comparison rather than quote an old figure. Docs: Fireworks docs. Not verified on October 14, 2026: price, limits, features.
Which one to pick for long audio
- Lowest cost and you are willing to split files yourself: Groq, especially its Batch API if you can wait up to a day.
- Single files longer than 4 hours: AssemblyAI (up to 10 hours per file).
- Speaker labels included, and the largest signup credit: Deepgram ($200).
- Huge monthly volumes: Salad's tiered pricing, or its Lite tier if you only need text.
- A back catalog of public recordings sent as a list of URLs, with speaker labels and SRT/VTT in one call: MyAi. Long files are split and stitched for you, and a webhook tells you when the batch is done.
- Confidential, medical, legal or regulated audio: not MyAi. Use a vendor whose data terms fit your obligations, or run Whisper on your own hardware.
How we checked
On October 14, 2026 we opened each vendor's pricing page and speech-to-text documentation and copied the figures as written. Per-minute prices were multiplied by 60. We did not run accuracy benchmarks for this article, so it makes no accuracy claims; accuracy depends on your audio, so test a few of your own files on two or three of these before you commit. Missing or unconfirmed details are marked "Not documented" or "Could not verify". If you spot a mistake, tell us and we will fix it.
FAQ
Which Whisper API is lowest priced per audio hour?
Of the list prices we checked on October 14, 2026, Groq is lowest: whisper-large-v3-turbo at $0.04 per audio hour, or $0.02 through its Batch API, which returns results within 24 hours to 7 days. Salad's Transcription API Lite ($0.08), Together AI's Whisper Large v3 ($0.09) and MyAi ($0.10) come next.
Which transcription APIs accept files longer than an hour?
AssemblyAI accepts up to 10 hours per file and 5 GB by URL. Deepgram accepts 2 GB per file but times out requests that take over 10 minutes to process. Together AI and MyAi accept up to 4 hours per file. Groq (25 MB or 100 MB) and OpenAI (25 MB) cap the file size, so long recordings must be compressed or split first.
Which ones include speaker labels?
Deepgram includes diarization for pre-recorded audio. AssemblyAI charges $0.02 per hour extra. Together AI and Salad offer it, and MyAi includes it. OpenAI offers it only on gpt-4o-transcribe-diarize. Groq's speech-to-text docs do not mention it.
Which ones return SRT or VTT captions directly?
AssemblyAI (SRT and VTT endpoints), MyAi (SRT, VTT, TXT, JSON), Replicate's openai/whisper (SRT or VTT output) and OpenAI's whisper-1. Salad returns SRT. Deepgram converts its JSON with separate caption libraries. Groq, Together AI and OpenAI's gpt-4o models return JSON or text.