Podrobný průvodce již brzy
Pracujeme na komplexním vzdělávacím průvodci pro Speech-to-Text API Cost Calculator. Brzy se vraťte pro podrobné vysvětlení, vzorce, příklady z praxe a odborné tipy.
The Speech-to-Text Cost Calculator estimates the expense of transcribing audio into text using AI services including OpenAI Whisper API ($0.006 per minute), Google Cloud Speech-to-Text ($0.016 per minute for standard models), AWS Transcribe ($0.024 per minute), Deepgram ($0.0043 per minute for Nova-2), and AssemblyAI ($0.0065 per minute). These services convert spoken language into written text with 90 to 98 percent accuracy depending on audio quality, language, and model selection. This calculator serves podcast production companies, call center analytics teams, legal transcription services, media companies captioning video content, and meeting productivity tools that generate automated transcripts. A 60-minute podcast episode costs $0.36 to transcribe with Whisper API, $0.96 with Google Speech, or $0.26 with Deepgram. At scale, these differences compound significantly: a call center transcribing 10,000 hours of calls per month would pay $3,600 with Whisper, $9,600 with Google, or $2,580 with Deepgram. Beyond basic transcription cost, the calculator also accounts for features that affect pricing: speaker diarization (identifying who said what), real-time versus batch processing (real-time typically costs 2 to 3 times more), specialized vocabulary models, and post-processing costs for formatting, punctuation insertion, and paragraph segmentation. Understanding the full cost stack helps teams choose the right service for their accuracy requirements and budget constraints.
Transcription Cost = Audio Duration in Minutes x Price per Minute. For batch processing 500 hours of audio per month on Whisper API: Cost = 500 x 60 x $0.006 = $180.00 per month. For real-time transcription on Google Cloud at 2x rate: Cost = 500 x 60 x $0.032 = $960.00.
- 1Determine your total audio duration to transcribe per month. Measure in minutes or hours and distinguish between pre-recorded (batch) and live (real-time) audio. Batch transcription is typically cheaper and can accept longer processing times. Real-time transcription delivers results as audio streams in but costs 1.5 to 3 times more due to the compute intensity of low-latency processing.
- 2Select your transcription service based on accuracy requirements, language support, and budget. Whisper API offers the best price-to-quality ratio for most languages at $0.006 per minute. Deepgram Nova-2 is the cheapest option at $0.0043 per minute with competitive accuracy. Google Cloud and AWS offer the broadest language support and integration with their respective cloud ecosystems. AssemblyAI provides the best speaker diarization and content analysis features.
- 3Configure transcription features that affect pricing. Speaker diarization (identifying different speakers) adds 10 to 30 percent to base transcription costs on most platforms. Punctuation and capitalization are included by default on modern services. Real-time streaming costs 2 to 3 times more than batch processing. Custom vocabulary or industry-specific models may have additional costs on some platforms.
- 4Calculate the base transcription cost by multiplying total audio minutes by the per-minute rate. For mixed real-time and batch workloads, calculate each separately: batch audio at the standard rate and real-time audio at the elevated rate. Sum the two for total monthly transcription expense.
- 5Add post-processing costs if applicable. Raw transcription output often needs formatting: paragraph breaks, speaker labels, timestamp insertion, filler word removal, and domain-specific corrections. Automated post-processing using an LLM (GPT-4o-mini) costs approximately $0.001 to $0.005 per minute of audio processed. Manual post-processing by a human editor costs $0.50 to $2.00 per audio minute depending on required accuracy.
- 6Factor in storage costs for audio files and transcripts. Audio files at 128 kbps consume approximately 1 MB per minute. Transcript text is negligible in size. Cloud storage at $0.02 per GB per month means 10,000 hours of audio (600 GB) costs $12 per month to store. If you need to retain audio for compliance, include this ongoing storage cost.
- 7Compare total cost against human transcription services. Professional human transcription costs $1.00 to $3.00 per audio minute for standard turnaround and $2.00 to $5.00 for rush delivery. AI transcription at $0.004 to $0.024 per minute is 40 to 750 times cheaper. Even with human quality review adding $0.25 to $0.50 per minute, AI-assisted transcription remains 50 to 85 percent cheaper than fully manual transcription.
40 episodes at 60 minutes each total 2,400 audio minutes. At $0.006 per minute, the monthly cost is $14.40. This replaces a human transcription service that would charge $2,400 to $7,200 per month for the same volume. The 99 percent cost reduction makes full transcription economically viable for every episode.
5,000 hours (300,000 minutes) of real-time call transcription with speaker diarization at $0.0059 per minute. Deepgram streaming pricing includes diarization. This enables real-time agent assistance and post-call analytics for compliance and quality assurance at a fraction of the cost of dedicated QA analysts.
30 depositions at 120 minutes each total 3,600 audio minutes. AI transcription costs $23.40. Human review at 15 minutes per audio hour (900 hours total review time) at $60/hr adds $450. Total is $473.40 versus $7,200 to $14,400 for fully human court reporter transcription.
200 videos at 15 minutes each total 3,000 audio minutes. Transcription at $0.016 per minute is $48.00, plus caption formatting at $0.002 per minute adds $6.00. ADA compliance captioning for 200 videos at $54 per month makes accessibility affordable for content creators of all sizes.
Podcast hosting platforms offer automatic transcription as a premium feature. A platform hosting 10,000 podcasts with an average of 4 episodes per month at 45 minutes each transcribes 1.8 million audio minutes monthly. Using Whisper API at $0.006 per minute, the monthly cost is $10,800. Charged at $5 per month to podcast creators as a premium feature, with 3,000 subscribers generating $15,000 in revenue, the feature is profitable while providing accessibility and SEO benefits.
Healthcare organizations transcribe physician-patient encounters for medical record documentation. A hospital network with 500 physicians averaging 20 patient encounters per day at 15 minutes each generates 150,000 audio minutes per month. Using a HIPAA-compliant service at $0.01 per minute costs $1,500 per month. Medical coders review transcripts in 5 minutes per encounter at $30 per hour, adding $25,000 monthly. Total cost of $26,500 replaces $75,000 per month in manual medical transcription.
Media companies caption video content for accessibility compliance and SEO. A streaming platform with 5,000 hours of new content per month uses Google Cloud Speech at $0.016 per minute, costing $4,800 monthly for transcription. Automated caption formatting adds $600. Human QA review at 10 minutes per content hour adds $8,333. Total captioning cost of $13,733 per month compared to $50,000 to $100,000 for manual captioning services.
Market research firms transcribe focus groups and customer interviews. A firm conducting 200 interviews per month at 45 minutes each uses AssemblyAI with speaker diarization at $0.0065 per minute, costing $58.50 per month for transcription. Researchers spend 10 minutes per interview reviewing and annotating transcripts at $75 per hour, adding $2,500. The $2,558.50 monthly cost enables rapid analysis that previously required dedicated transcription staff at $8,000 per month.
For HIPAA-compliant medical transcription, not all services are eligible.
Google Cloud Speech, AWS Transcribe, and Azure Speech offer HIPAA-compliant configurations with Business Associate Agreements. OpenAI Whisper API and Deepgram offer enterprise agreements with compliance certifications. Self-hosted Whisper on HIPAA-compliant infrastructure provides the most control but requires dedicated security and compliance expertise. Medical transcription also benefits from custom vocabulary models trained on medical terminology.
For transcribing audio in noisy environments (construction sites, restaurants,
For transcribing audio in noisy environments (construction sites, restaurants, outdoor events), accuracy can drop to 60 to 75 percent even with the best models. Pre-processing with noise reduction tools like RNNoise or DeepFilterNet can improve accuracy by 10 to 20 percentage points at negligible computational cost. For extremely noisy audio, a two-pass approach (noise reduction followed by transcription) is more accurate and ultimately cheaper than relying on human correction of low-quality transcripts.
For archival transcription of historical audio recordings (analog recordings,
For archival transcription of historical audio recordings (analog recordings, old phone systems, degraded tape), audio quality issues compound with the technology limitations of older models. These recordings may have low sample rates (8kHz telephone), significant background noise, and outdated vocabulary. Specialized preprocessing to upsample and denoise the audio, combined with fine-tuned models for the specific audio quality, can improve accuracy from 50 to 60 percent (standard models) to 80 to 90 percent, at an additional preprocessing cost of $0.002 to $0.01 per minute.
| Service | Batch Price/Min | Streaming Price/Min | Diarization | Languages | Free Tier |
|---|---|---|---|---|---|
| OpenAI Whisper API | $0.006 | N/A | No | 99+ | No |
| Deepgram Nova-2 | $0.0043 | $0.0059 | Yes ($0.0049) | 36+ | 12,000 min |
| AssemblyAI | $0.0065 | $0.0085 | Yes (included) | 20+ | 100 hrs |
| Google Cloud Speech | $0.016 | $0.032 | Yes ($0.02) | 125+ | 60 min/mo |
| AWS Transcribe | $0.024 | $0.024 | Yes (included) | 100+ | 60 min/mo (12 mo) |
| Azure Speech | $0.016 | $0.016 | Yes ($0.02) | 100+ | 5 hrs/mo |
Which speech-to-text service is most accurate?
For English, OpenAI Whisper and Deepgram Nova-2 achieve the highest accuracy at 95 to 98 percent word error rate on clean audio. Google Cloud Speech and AWS Transcribe are comparable at 93 to 97 percent. For specialized domains (medical, legal, financial), custom vocabulary models on Google Cloud or AWS can improve accuracy by 3 to 5 percentage points. Accuracy drops 5 to 15 percentage points for noisy audio, heavy accents, or multilingual content.
How does Whisper API compare to self-hosted Whisper?
OpenAI Whisper API at $0.006 per minute is a managed service with no infrastructure to manage. Self-hosted Whisper on a cloud GPU (A10G at $1.10/hr) processes audio at approximately 10 to 30 times real-time speed, costing $0.001 to $0.002 per minute at full utilization. Self-hosting is 3 to 6 times cheaper but requires GPU management, scaling, and monitoring. The break-even is approximately 5,000 to 10,000 audio minutes per month.
What is the difference between batch and real-time transcription?
Batch transcription processes pre-recorded audio files and returns results in minutes to hours. Real-time (streaming) transcription processes audio as it is captured and returns words within 200 to 500ms of being spoken. Batch is 1.5 to 3 times cheaper and is suitable for recorded content. Real-time is essential for live captioning, meeting transcription, and call center agent assistance. Most providers offer both modes.
How accurate is AI transcription for meetings with multiple speakers?
Modern services achieve 90 to 95 percent accuracy for meetings with clear turn-taking between 2 to 4 speakers. Accuracy drops to 80 to 90 percent with overlapping speech, more than 6 speakers, or poor microphone quality. Speaker diarization (identifying who said what) is 85 to 95 percent accurate for well-separated speakers but can fall below 80 percent in chaotic meeting environments. Using high-quality microphones and recording setups significantly improves results.
Can I transcribe in languages other than English?
Yes, all major services support multiple languages. Whisper supports 99+ languages with varying accuracy. Google Cloud supports 125+ languages. Accuracy for non-English languages is typically 3 to 10 percentage points lower than English. For tonal languages like Mandarin and Thai, specialized models offer better accuracy. Cost per minute is generally the same regardless of language, though some services charge a premium for less common languages.
Pro Tip
For maximum cost efficiency on large transcription workloads, self-host Whisper on a cloud GPU with automatic scaling. An A10G GPU at $1.10 per hour transcribes audio at approximately 15 times real-time speed, costing about $0.001 per minute. Set up an auto-scaling queue that spins up GPUs when audio files are submitted and shuts them down when the queue is empty. This approach saves 70 to 85 percent versus the Whisper API while handling variable workloads efficiently.
Did you know?
OpenAI Whisper was trained on 680,000 hours of multilingual audio, the equivalent of listening non-stop for 77 years. Despite being released as an open-source model in 2022, the Whisper API remains one of the most popular transcription services because the convenience of API access outweighs the cost savings of self-hosting for most organizations.