Who whisper-asr-webservice is for#
Backend developers building transcription pipelines
Whisper ASR Webservice exposes a REST API with Swagger documentation, making it straightforward to integrate into existing services via HTTP. Developers can POST an audio file and receive JSON with word-level timestamps or ready-to-use SRT subtitles without writing any ASR code.
Skip if:
Skip if you need real-time or streaming transcription. This tool processes complete audio files and is not suited for live captioning workflows where partial results must arrive within seconds.
Teams processing meeting recordings or podcasts
Speaker diarization via the WhisperX engine identifies individual speakers across a recording, making it well-suited for meeting transcript generation, podcast chapter creation, and interview processing. Output in SRT or JSON maps directly to video editing or CMS workflows.
Skip if:
Skip if audio quality is poor or speakers frequently overlap. WhisperX speaker separation performs best on clean recordings with distinct speakers.
Organizations with audio data privacy requirements
Self-hosting keeps audio on infrastructure you control. For healthcare, legal, or financial teams where sending recordings to a third-party API creates compliance risk, running Whisper locally removes the data-transfer concern entirely.
Skip if:
Skip if your compliance requirement extends to the model weights themselves. Whisper is trained on data from the internet; some regulated environments require audited, approved models.
Developers building subtitle generation workflows
The API returns SRT and VTT files directly, with no post-processing needed to create subtitle tracks for videos. Combined with word-level timestamps, it covers the full subtitle pipeline from raw audio to a time-coded subtitle file.
Skip if:
Skip if you need subtitles with real-time sync during live video streaming. Batch processing means subtitles are available after the full file is processed, not during playback.
The problem it solves#
Transcribing audio at any meaningful scale with commercial APIs means paying per minute of audio processed. Deepgram, AssemblyAI, Amazon Transcribe, and the OpenAI Whisper API all bill per second or per request, and costs add up quickly for teams processing meeting recordings, podcasts, or video archives. A team transcribing 100 hours of audio per month can spend several hundred dollars before accounting for development costs or usage spikes.
Beyond cost, sending audio to a third-party API creates data exposure concerns. Sensitive conversations, internal meetings, and customer calls move outside your control the moment you POST to a cloud endpoint. Many compliance environments (healthcare, legal, finance) prohibit this entirely. Even teams without formal compliance requirements may want their audio data to stay on infrastructure they own.
How it solves it#
Multiple ASR Engine Support
Choose between three ASR backends: the original openai/whisper for broad model support, SYSTRAN/faster-whisper for lower memory use and faster inference, and whisperX for speaker diarization. Switch engines via the ASR_ENGINE environment variable.
Output in Five Transcript Formats
Returns transcripts as plain text, JSON with timestamps, SRT subtitles, VTT subtitles, or TSV. The SRT and VTT formats attach directly to video files; the JSON format includes word-level timestamp data for downstream processing.
Word-Level Timestamps
Every transcribed word carries a start and end timestamp in the JSON output. This enables time-synced search, precise clip extraction, and highlight generation without a second pass over the audio.
Voice Activity Detection Filtering
VAD filtering removes silence and non-speech segments before transcription. This reduces processing time and improves accuracy on recordings with long pauses, hold music, or background noise between speech segments.
Speaker Diarization via WhisperX
When running the WhisperX engine, the API adds speaker labels to the transcript, identifying who said what across a multi-speaker recording. This is useful for meeting transcripts, interview processing, and podcast show notes.
GPU Acceleration and CPU Fallback
A GPU image with CUDA support is available for teams that want faster inference on larger Whisper models. The standard CPU image runs on any machine without a GPU, making both development and low-volume production deployable without specialized hardware.
Strengths and trade-offs#
Strengths
- No Per-Minute BillingUnlike Deepgram, AssemblyAI, or the OpenAI Whisper API, Whisper ASR Webservice charges nothing per transcription. After the infrastructure cost of running the container, every additional hour of audio processed costs the same: zero. Teams with high-volume transcription workloads recover server costs quickly.
- Full Audio Data OwnershipAudio never leaves your infrastructure. Every file uploaded to the API is processed locally and the bytes do not transit to any third-party service. This makes deployment viable in healthcare, legal, and other regulated environments where external data transfer is restricted.
- MIT License with No Usage RestrictionsThe MIT license means you can run this commercially, modify the code, and ship it as part of a product without any licensing restrictions. This contrasts with proprietary APIs that prohibit resale or modification.
- FFmpeg Integration for Broad Format SupportFFmpeg handles audio decoding, so the API accepts MP3, MP4, WAV, OGG, M4A, and other common formats without the caller needing to pre-convert to a specific codec. Almost any audio or video file works without preprocessing.
Trade-offs
- -Self-Hosting Requires Server ResourcesRunning Whisper models requires meaningful CPU or GPU resources. The base model fits on lightweight hardware, but the large-v3 model needs at least 10GB of VRAM on GPU or substantial RAM on CPU. Teams on shared hosting or small VPS instances may find only the tiny and base models practical.
- -122 Open IssuesThe repository carries 122 open GitHub issues. While the project is actively maintained with a push in August 2026, a large open issue count means some edge cases and integrations may have known bugs without immediate fixes.
- -Higher Latency Than Managed Cloud APIsWhisper inference is slower than the real-time transcription offered by Deepgram or AssemblyAI's streaming endpoints. This tool processes audio files in batch; it is not suitable for live captioning or low-latency transcription pipelines where results need to arrive within seconds.
whisper-asr-webservice vs alternatives#
Whisper ASR Webservice vs OpenAI Whisper API
The OpenAI Whisper API and Whisper ASR Webservice use the same underlying model family. The deciding factor is where inference runs: on OpenAI's servers or yours.
| Feature | Whisper ASR Webservice | OpenAI Whisper API |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes | No |
| Pricing | Free (server cost) | Per-minute fee |
| Speaker diarization | Yes (WhisperX engine) | No |
| Output formats | Text, JSON, SRT, VTT, TSV | JSON only |
| GPU acceleration | Yes | Managed |
Whisper ASR Webservice is the better choice when audio data cannot leave your infrastructure, when transcription volume makes per-minute billing expensive, or when you need output formats the API does not provide (SRT, VTT, word-level JSON). The OpenAI Whisper API is easier to integrate for small-volume use cases: no infrastructure to manage, no model downloads, and it works from an API key in minutes.
Whisper ASR Webservice vs Deepgram
Deepgram is a real-time capable, proprietary speech-to-text API with its own nova and enhanced models. The two tools serve meaningfully different use cases.
| Feature | Whisper ASR Webservice | Deepgram |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes | No |
| Real-time streaming | No | Yes |
| Speaker diarization | Yes | Yes |
| Pricing | Free (server cost) | Per-minute fee |
Whisper ASR Webservice wins on cost control and data privacy. Deepgram wins on latency: its streaming endpoint delivers partial transcripts as audio arrives, making it the right tool for live captioning, voice agents, and anything that cannot wait for a complete file to be processed. If your use case is batch transcription of recorded audio and data must stay on your infrastructure, Whisper ASR Webservice is the stronger choice.
Whisper ASR Webservice vs AssemblyAI
AssemblyAI is a managed transcription API with add-on features like sentiment analysis, chapter detection, and content moderation. Whisper ASR Webservice covers the transcription core but not those higher-level features.
| Feature | Whisper ASR Webservice | AssemblyAI |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes | No |
| Speaker diarization | Yes | Yes |
| Sentiment analysis | No | Yes |
| Content moderation | No | Yes |
| Pricing | Free (server cost) | Per-minute fee |
Choose Whisper ASR Webservice when transcription is the end goal and cost or privacy requirements make a cloud API impractical. Choose AssemblyAI when you need value-added intelligence on top of the transcript (sentiment, chapters, content safety) and can accept per-minute billing and third-party data handling.
Install and self-host#
Deploy using Docker; the standard image runs on CPU and the latest-gpu tag requires CUDA-compatible hardware.
```bash
docker run -d -p 9000:9000 -e ASR_MODEL=base -e ASR_ENGINE=openai_whisper onerahmet/openai-whisper-asr-webservice:latest
docker run -d --gpus all -p 9000:9000 -e ASR_MODEL=base -e ASR_ENGINE=openai_whisper onerahmet/openai-whisper-asr-webservice:latest-gpu
```What it's built on#
- Languages
- Python
- Frameworks
- FastAPI
- Infrastructure
- Docker
FAQ#
Is Whisper ASR Webservice free to use?
Yes. The project is MIT licensed and free to self-host. You pay only for the server or cloud compute that runs the Docker container. There is no per-transcription fee, no API key cost, and no usage cap imposed by the software.
What audio formats does Whisper ASR Webservice support?
The service uses FFmpeg for audio decoding, so it accepts most common audio and video formats including MP3, MP4, WAV, OGG, and M4A. Any format that FFmpeg can read can be passed to the API without pre-conversion.
Does Whisper ASR Webservice support multiple languages?
Yes. The underlying Whisper model is trained on multilingual audio and supports transcription and translation across dozens of languages. The API also supports language identification, so the model can detect the spoken language automatically.
How does it compare to the paid OpenAI Whisper API?
Both use the same Whisper model family. The key difference is infrastructure: the OpenAI Whisper API is a managed cloud endpoint that charges per audio minute; Whisper ASR Webservice runs the model on your own server with no per-request fee. The self-hosted version also adds speaker diarization and multiple output formats not available through the OpenAI API.
Does it support GPU acceleration?
Yes. A GPU-enabled Docker image is available for systems with CUDA-compatible hardware. The GPU build significantly reduces inference time for larger Whisper models like medium and large-v3. The standard CPU image runs on any machine without requiring a GPU.
Similar open-source tools#
RealtimeSTT
Real-time speech-to-text library with VAD and wake words
Switchyard
LLM proxy with API translation and multi-backend routing
Embabel
Agentic AI framework for the JVM
Agent-Reach
Give agents local web and social-source access
Remodex
Control Codex agents from your iPhone, open source and free
dograh
Open source voice agent builder with telephony and BYOK AI support