The complete unit economics guide for AI voice bots, call center transcription, and meeting note agents. Compare latency, word error rates (WER), and volume discounts.
| Provider & Model | Cost / Minute | Cost / Hour | 1,000 Hours Cost | Streaming Latency | Word Error Rate (WER) | Best For |
|---|---|---|---|---|---|---|
| Groq Whisper Large v3 | $0.0018 | $0.111 | $111 | Batch only (250x RT) | 7.2% | Massive batch audio archives |
| Deepgram Nova-3 | $0.0043 | $0.258 | $258 | ~220ms | 6.84% | Real-time voice bots, LiveKit |
| Deepgram Nova-2 | $0.0043 | $0.258 | $258 | ~260ms | 7.80% | Legacy telephony pipelines |
| OpenAI Whisper Large v3 | $0.0060 | $0.360 | $360 | Batch only | 7.4% | Multilingual translation |
| AssemblyAI Best | $0.0065 | $0.390 | $390 | ~350ms | 7.1% | PII redaction, call analytics |
| Gladia Audio Intelligence | $0.0090 | $0.540 | $540 | ~300ms | 7.5% | Multi-speaker overlap meetings |
| Google Cloud Speech-to-Text | $0.0160 | $0.960 | $960 | ~400ms | 8.4% | Google Workspace integrations |
For conversational AI voice bots (e.g. AI receptionists or support agents), Time to First Transcript Token (TTFT) is everything. If your user says "hello" and transcription takes 800ms, your entire conversational turn will feel sluggish and unnatural. In this scenario, Deepgram Nova-3 via streaming WebSockets is the only viable choice under $0.005/min.
Conversely, if your workload is transcribing recorded customer service calls or user voicemails after the fact, streaming low-latency is completely irrelevant. Routing recorded files to Groq Whisper Large v3 at $0.0018/min reduces your transcription expenses by nearly 60% with zero quality degradation.
Compare monthly transcription invoices across DeepInfra Whisper ($0.09/hr), Deepgram Nova-3 ($0.258/hr), Groq Whisper ($0.111/hr), and OpenAI Whisper ($0.360/hr).
Python script streaming live caller audio over WebSockets with partial transcript callbacks under 250ms.
import os
import asyncio
from deepgram import DeepgramClient, LiveTranscriptionEvents, LiveOptions
async def stream_audio_to_deepgram(audio_stream):
"""Stream live telephony audio to Deepgram Nova-3 for sub-300ms transcription."""
deepgram = DeepgramClient(os.environ["DEEPGRAM_API_KEY"])
connection = deepgram.listen.websocket.v("1")
def on_transcript(self, result, **kwargs):
sentence = result.channel.alternatives[0].transcript
if sentence:
print(f"[Live Partial Transcript]: {sentence}")
connection.on(LiveTranscriptionEvents.Transcript, on_transcript)
options = LiveOptions(model="nova-3", language="en-US", punctuate=True, diarize=True)
await connection.start(options)
for chunk in audio_stream:
await connection.send(chunk)
await connection.finish()Common engineering traps that cause latency spikes and unexpected charges.
OpenAI Whisper frequently hallucinates repetitive phantom phrases ('Thank you for watching!') during long audio pauses. Use Voice Activity Detection (VAD) to strip silence.
Features like speaker identification (who spoke when) add $0.0014 to $0.0030/min surcharges across Deepgram and AssemblyAI. Verify feature flags before batch jobs.
OpenAI and Groq Whisper endpoints reject files over 25 MB. Multi-hour podcast recordings must be chunked with ffmpeg into 10-minute MP3 segments.
Sending raw uncompressed WAV audio over mobile networks causes bandwidth choke. Encode audio to 16kHz mono Opus or 64kbps MP3 to minimize bandwidth bills.
Explore complementary benchmarks across speech synthesis and conversational telephony.
Compare sub-300ms WebSocket streaming against batch transcription.
Compare state-space Sonic voice vs ElevenLabs Multilingual v2.
Evaluate CPaaS voice minutes and SMS routing economics.
Rankings of the lowest cost generative voice APIs.
Speech-to-Text (STT) infrastructure requires evaluating APIs on four interdependent metrics: Word Error Rate (WER) under background noise, streaming latency (TTFB), speaker diarization surcharges, and per-minute pricing. Legacy hyperscalers (Google Cloud Speech-to-Text and AWS Transcribe) charge between $0.016 and $0.024 per minute ($0.96 to $1.44/hour). Specialized speech AI providers like Deepgram Nova-2 ($0.0043/min, $0.258/hour) and Whisper Large-v3 hosted on Groq LPUs or DeepInfra offer 70% to 80% lower operational costs with superior handling of conversational interruptions and industry jargon.
| Provider & Model | Base Price / Min | Hourly Rate | Streaming Latency | WER (Clean / Noisy) | Monthly Total (2,500 Hrs) |
|---|---|---|---|---|---|
| Deepgram Nova-2 | $0.0043 | $0.258 | 240ms | 6.8% / 11.2% | $645.00 |
| Whisper Large-v3 (Groq LPU) | $0.0018 (Batch) | $0.111 | Chunked (Non-stream) | 7.2% / 12.0% | $277.50 |
| AssemblyAI (Conformer-2) | $0.0065 | $0.390 | 450ms | 7.1% / 11.8% | $975.00 |
| Google Cloud STT v2 | $0.0160 | $0.960 | 380ms | 8.4% / 13.9% | $2,400.00 |
import asyncio
import websockets
import json
# Stream live audio frames into Deepgram with endpointing and interim transcripts
async def stream_audio_to_deepgram(audio_queue):
url = "wss://api.deepgram.com/v1/listen?model=nova-2&smart_format=true&interim_results=true&endpointing=300"
headers = {"Authorization": f"Token {os.environ['DEEPGRAM_API_KEY']}"}
async with websockets.connect(url, extra_headers=headers) as ws:
async def send_audio():
while True:
chunk = await audio_queue.get()
if chunk is None: break
await ws.send(chunk)
async def receive_transcripts():
async for msg in ws:
res = json.loads(msg)
transcript = res.get('channel', {}).get('alternatives', [{}])[0].get('transcript', '')
if transcript:
print("STT:", transcript)
await asyncio.gather(send_audio(), receive_transcripts())Whisper was architected for batch audio and processes files in 30-second fixed chunks. While specialized providers like Groq can transcribe a 30-second chunk in under 300ms, streaming word-by-word interim results requires streaming architectures like Deepgram Nova-2 or Conformer models.
Raw linear PCM (16-bit, 16kHz, mono) delivers the lowest latency because audio frames require zero decode buffering on the server. For mobile cellular streams where bandwidth is constrained, Opus at 24 kbps provides superior audio fidelity with negligible decompression overhead.