Mano, tentando fazer uma aplicação rápida aqui pra ajudar quem for checar. Usando o Deepgram e o Perplexity pra transcrever o debate, checar os dados, e o voluntário só aprovar ou não. Testando com o debate de semana passada. Se der certo te falo
brandCooling
Deepgram on X
The biggest theme is Deepgram's own figure — roughly 60–70% more audio per dollar on AWS Trainium — repeated more than tested. Developers praise Flux's speed and turn control; complaints are early cutoffs, unverified claims and an iOS Nova-3 retry loop, with Voxtral 3 priced against it.
First written 27 Oct · what changed · from 49 posts by 44 people
Tracking posts that say “deepgram”
Track this for your marketPosts, last 24h
12
−50%vs. previous 24h
Posts, last 7 days
146
69 in the last 72h
Velocity
21/day
7-day average
Momentum
−42%
Last 24h vs. a normal day this week
Posts per day, last 30 days
as of Nov 3, 12:00 UTCWhat changed
as of 27 Oct · checked once a monthFirst written 27 Oct. When what people say changes, it shows here.
Tone· 31 posts from 28 people· written 27 Oct
How do people feel about Deepgram?
The biggest theme is Deepgram's own cost number: 11 people repeat its claim of roughly 60–70% more audio processed per dollar on AWS Trainium. Praise for it (7 people) and people showing products built on it (7 people) are level behind that.
The Trainium theme is largely one event repeated: Amazon's TorchNeuron launch on 20 October, where Deepgram's CTO was the customer quoted in the announcement. Posts carry the figure differently — about 60% more audio per dollar, some 70% — and treat it as a procurement argument rather than their own product experience. It is Deepgram's claim, repeated by others, and one post notes the port still took months of work alongside AWS engineers.
The praise is about speed and control. Flux users single out end-of-turn reports with probabilities that they tune their agent against, and say accuracy and responsiveness are dialled in; one shows Nova-3 catching a company's own strategy name all 14 times only after key terms were added, having written "Monroe" before. Products built on it range from a restaurant-booking call agent and a bot costing about three cents a call to a support desk demo and a debate-transcription app that travelled furthest of these posts.
A minority of 3 people call it decent rather than best, rating ElevenLabs ahead on results, and 2 are pleased how far the $200 signup credit goes.
- Cheaper audio on Trainium11
- Praises it8
- Used in real products7
- Decent but not the best3
- Pleased by $200 credits2
See 5 of 31 postsHide the posts
Dilini bilmediğiniz bir memlekette sizin adınıza restoranı arayıp rezervasyon yapan bir agentınız olsa nasıl olurdu? Twilio , deepgram, elevenlabs, deepseek alt yapılarını kullanarak bir test yaptım. Hem aramayı dinleyebiliyorsunuz, hem alt yazıları takip edebiliyorsunuz. Sonunda da size bir özet veriyor. Aşağıda philip adına 22 ekim 19:00 için 3 kişilik bir rezervasyon yapmaya çalışan ilk denemem var. Siz kullanır mıydınız?
之前聊英伟达护城河的时候，核心逻辑是CUDA绑住开发者，换芯片意味着代码重写，没人付得起这个成本。亚马逊昨天推的TorchNeuron，拆的就是这一层。 工具本身不复杂：开发者写好的PyTorch代码，不用重写，改一处设备配置，就能在亚马逊自研的Trainium芯片上跑起来。亚马逊工程师拿同一份训练脚本分别在传统芯片和Trainium上跑，模型学到的内容完全一致，速度也相同。官方说法很直白：如果你懂PyTorch，你就懂Trainium。 真正指向CUDA软肋的地方在于，过去把负载从英伟达迁走，通常要花几周甚至几个月做适配和重新验证，这笔工程成本比硅片价格更能把人摁在英伟达身上。TorchNeuron想砍的就是这个环节。 目前已经有超过30家早期客户在用，覆盖机器人、世界模型和语音AI，官方说转用Trainium能带来显著的成本节约。语音AI公司Deepgram给了一个更具体的数据：在它评估的基准里，Trainium每美元处理的输入音频秒数比现有GPU基线多出大约60%。 不过这只是第一步。首版TorchNeuron明确不支持动态和符号形状，也就是模型执行过程中数据维度会变的场景，这部分今天还跑不了。CUDA的护城河不只是“能不能跑”，还有跑了十年攒下来的库、调试工具和踩过的坑。拆掉迁移成本这道门槛，不等于拆掉了整座城墙。 而且亚马逊不是一个人在拆。谷歌也在做TorchTPU，目标同样是让PyTorch代码用最小改动迁到TPU上。微软上周也表态要把大部分AI服务迁到自研Maia芯片上。 所以亚马逊这一步，短期改变不了英伟达在训练端的地位，但它把“换芯片到底有多贵”这个问题的答案往下压了一截。当换芯片从伤筋动骨变成改行参数，买家手里的议价空间就多了一点。护城河还在，只是墙根底下被人凿了一个洞，洞会不会变大，看接下来有多少人愿意从那个洞钻过去。 $NVDA $AMZN $GOOGL 免责声明：本文纯属个人观点，不构成任何投资建议。
Yes, my Spectra system is using Deepgram Flux for the STT! I love the end of turn reports with probabilities, I pair those with Jev to know when the agent should respond. I've got it dialed in so the accuracy is very high and responsiveness is ultra quick. Love Deepgram
I gave @DeepgramAI Nova-3 key terms and I pulled automatically from each company's 10-K and ran it on earnings calls. I counted 14 mentions of Monro's own strategy name, Monro Forward, in one call. I saw Nova-3 get it right only once with no keyterms, since it kept writing Monroe. I added the 10-K keyterms and watched it get all 14. #deepgram #AI
5 of the 31 posts behind this answer, one per person. The other 26, and every new one as it’s posted, come with tracking. Get early access
Pain points· 9 posts from 8 people· written 27 Oct
What do people complain about with Deepgram?
Complaints split evenly at the top: turn detection and cutoffs and doubt over the Trainium cost numbers are level at 3 people each, with an iOS retry loop just behind at 2 people across 3 posts.
On turn detection, the complaint is Flux ending a turn early. One developer building a phone voice agent says Flux ended their turn inside a mid-sentence pause four times in one call, with them resuming 0.46 to 1.3 seconds after each cut, and shipped a single threshold per call type in response. The other side of the tradeoff is named too: Flux's end-of-turn settings raise the confidence needed before a turn is called finished, which means fewer mid-sentence interruptions but slower replies, and the wait for a transcript to finalise after a pause sets the stable window.
The Trainium doubt is about whose number it is. Posters call the 70% more audio per dollar a company claim rather than an independent benchmark, note the offline batch range is much wider than that headline, and say to test it on your own workload; one says the efficiency gain needed custom kernels and serving work beyond porting the model.
The retry loop is specific: Nova-3 transcription on an iPhone 17 Pro running iOS 27.0.1, where every hourly slot tested returns straight to a retry countdown at 15:00 and 60:00, with recordings not processing and the cause not yet confirmed.
- Turn detection and cutoffs3
- Trainium claims unverified3
- iOS transcription retry loop3
See 5 of 9 postsHide the posts
building aria, a voice ai you call on the phone. shipped today: one end-of-turn threshold per call type. on one call deepgram flux ended my turn inside a mid-sentence pause 4 times, and i resumed 0.46 to 1.3s after every cut.
Good to see interruption measured. The TTS tail is one stage of barge-in, though. The full path is detecting the user’s speech, cancelling the LLM stream, stopping TTS, and flushing the audio already buffered on the client or telephony leg. Detection and buffer flush often cost more than the TTS stop itself. The methodology should also state what counts as a “stop request”: closing the socket, sending a Clear/flush message, or a cancel. Detection is tunable on the STT side. Deepgram Flux exposes three end-of-turn settings: - eot_threshold: the confidence Flux needs before calling the turn finished. Higher means fewer cases of the agent jumping in mid-sentence, but slower replies. - eager_eot_threshold: fires EagerEndOfTurn early, so the LLM can start drafting before the turn is confirmed. If the user keeps talking, TurnResumed tells you to throw that draft away. - eot_timeout_ms: forces end of turn after a stretch of silence, so a trailing “umm” doesn’t leave the agent waiting forever. On the interruption side, StartOfTurn fires when the user barges in. That’s your trigger to cancel the LLM, stop TTS, and clear the playback buffer. Each of these settings trades speed against false cutoffs, so the same TTS model can feel very different depending on how the pipeline is tuned. End-to-end barge-in latency is the number customers feel.
Amazon 10 月 19 日宣布 TorchNeuron，10 月 20 日在 PyTorch Conference 以 Apache 2.0 开源公开 beta。它支持 Trainium2／Trainium3，既有 PyTorch 程序改一行设定就能换芯片跑，官方称已有 30 多家早期客户在试。 这刀砍的是 CUDA 的生态锁，不是芯片规格。能不能成，先看社区接不接受 AWS 的上游改动，再看独立跑分；Deepgram 说每美元多处理 70% 音频是厂商口径的区间上限，实际离线批次是 11%–70%，待独立验证。 来源：link
The money pitch is AWS's own example: Deepgram processes 70% more audio per dollar after moving to Trainium. That's a company claim, not an independent benchmark. Test it on your own workload.
That Deepgram case confirms the efficiency gain required custom kernels and serving work beyond model porting.
5 of the 9 posts behind this answer, one per person. The other 4, and every new one as it’s posted, come with tracking. Get early access
Alternatives· 11 posts· written 27 Oct
What are the alternatives to Deepgram?
The largest group here, 4 people, weighs Voxtral 3 against Deepgram, mainly on price. ElevenLabs, Sarvam and local models follow with 2, 2 and 2 people each.
The Voxtral posts are about a price list, not a test drive. Mistral's launch figures are repeated — transcription at $0.0035 a minute and text-to-speech at $13 per million characters — with the claim that this undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. One post flags that three of the four TTS charts are labelled internal evals, and another notes Voxtral can run on the customer's own infrastructure. It is priced against Deepgram rather than measured against it.
ElevenLabs is where people go for output quality: it is called the better result and "the goat" for final voice, with Deepgram kept for speech-to-text, and one poster says none of the models tried, Deepgram included, gave conclusive results in French.
Sarvam is the India case — a live AI interview product running on Deepgram is considering switching because Sarvam is made for Indians, and a poster puts Saaras V4 at less than half what Deepgram or Azure charge per hour.
Local models are weighed for sovereignty, but one builder running Flux says nothing local has beaten it on latency yet. Whisper appears once, as a fallback after credits run out.
- Voxtral4
- ElevenLabs2
- Sarvam2
- Local models2
- Whisper1
See 5 of 11 postsHide the posts
Mistral released Voxtral 3 on October 20. It is a set of speech models plus a voice agent SDK built on LiveKit. On transcription, Mistral says Voxtral Small Transcribe 3 outperforms Whisper and Scribe v2. It says the model also beats Gemini 3.5 Transcribe. The realtime model is listed at sub-500ms latency. On speech output, Voxtral Mini TTS 3 costs $13 per 1M characters. Mistral says that is less than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. Three of its four text-to-speech charts are labeled internal evals. The SDK joins listening, reasoning and speaking in one streaming pipeline. Mistral says each turn completes in under one second. The price list is where I'd look. Mini Transcribe 3 in batch is $0.0035 a minute. Small Transcribe 3 is $0.0040 a minute. My read: Small is the model Mistral says beats Whisper and Scribe v2. It costs about 14% more than Mini. Over an hour of audio that is 21 cents against 24 cents. If Mistral's claims hold up on your own audio, the price gap alone is a weak reason to pick Mini for batch work. link
🚨 Mistral تطلق Voxtral 3: ذكاء اصطناعي صوتي يتكلم العربية الفصحى واللهجة الإماراتية، ويفرّغ الكلام بـ 0.0035$ فقط للدقيقة! 🤯 بناء وكيل صوتي حقيقي ما زال صعباً: عليه أن يسمع بدقة، ويفهم ما تقصده، ويتكلم بمشاعر طبيعية، وينفذ المهام.. وكل ذلك في نفس اللحظة. Mistral جمعت الأربعة في خط واحد. 🔥 ماذا في Voxtral 3: • تفريغ صوتي في 50+ لغة: نموذج Voxtral Small Transcribe 3 يتفوق على Whisper و Scribe v2 و Gemini 3.5 Transcribe حسب Mistral، ويميّز من قال ماذا. • يعرف متى انتهيت من الكلام ليرد الوكيل بلا تأخير، ونسخة Realtime بزمن أقل من 500ms. • صوت يقرأ المشاعر والإيقاع من سياق المحادثة في 35+ لغة و50+ لهجة، أو تتحكم فيه سطراً بسطر بوسوم مثل [emotion=sad][pace=slow]. • Pro Voice: استنساخ صوتك بدقة عالية من 30 دقيقة تسجيل فقط، في كل اللغات المدعومة. 📊 الأسعار: • التفريغ: من 0.0035$ للدقيقة. • توليد الصوت: 13$ لكل مليون حرف، أرخص من OpenAI TTS-1 و Deepgram و ElevenLabs Flash و Gemini Flash حسب Mistral. • وكيل صوتي كامل مع Mistral Small 4: فقط 0.04$ للدقيقة. 💡 كيف يعمل الوكيل الصوتي؟ SDK مبني على LiveKit: يسمع، ثم يفكر ويستدعي الأدوات، ثم يرد.. كل دور في المحادثة يكتمل في أقل من ثانية، ويتوقف عن الكلام فوراً إذا قاطعته. يستقبل المكالمات ويجريها عبر الهاتف (SIP)، ويعمل داخل المتصفح والتطبيقات، ويمكنك تشغيله على سيرفراتك الخاصة. وكيل يرد على عملائك بالعربية ويحجز ويتحقق من الفواتير بـ 4 سنتات للدقيقة.. خدمة العملاء لن تبقى كما هي! 🔥 احفظ التغريدة 🧵👇 الرابط في أول رد
$13 per 1M characters is the price of Mistral's new Voxtral 3 text-to-speech. Mistral says that is less than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. It runs on Mistral's EU-hosted API or on the customer's own infrastructure.
Voice AI for Indian consumers just got a lot cheaper. Sarvam's Saaras V4 covers all 22 scheduled Indian languages plus English for Rs 30 an hour, less than half of what Deepgram or Azure charge.
Currently our AI interview is being used by Indian students and we are using deepgram for stt ,we are thinking to switch it to sarvam as it is especially made for Indians. What are your views on it?
5 of the 11 posts behind this answer, one per person. The other 6, and every new one as it’s posted, come with tracking. Get early access
Wishes· 1 post· written 27 Oct
What do people want from Deepgram?
One person asks how much tighter Deepgram's timing and accuracy might get — 1 person, so there is no pattern here yet.
The only post filed here is a question, not a request: a developer replying to a Deepgram thread says they are curious how much tighter the timing and accuracy might get. Nothing in the section asks for a feature, a language, a price or a platform.
- Tighter timing and accuracy1
In their words
How do people talk about Deepgram?
The phrases people use most, counted once per person.
- 1“per dollar”6
- 2“audio per dollar”4
- 3“voice ai”4
- 4“custom kernels”3
- 5“gpu setup”3
- 6“end of turn”2
- 7“tts model”2
Posts are shown the way X shows them. One of them yours? Here’s why it’s shown, and how to take it off.
Also in Voice AI models and APIs
- ElevenLabs+2%The largest group, 46 of 124 people, talks about finished work — voiceovers, short films, games, receptionists — with the voices praised as hard to tell from a real person.392 posts / 7d
- Cartesia+2%The largest group of people post about what they built with Cartesia (7 people) — voice agents, nightly voice-over, cloned-voice pages — and 5 praise the voices themselves.268 posts / 7d
- AssemblyAI−44%Talk about AssemblyAI is builders showing what they made with it — dictation tools, call agents, hackathon demos — the largest group at 9 people, with short praise next.151 posts / 7d
- Speechify−36%Speechify's talk is mostly practical praise from people who listen to PDFs, books and textbooks instead of reading them, with price as the recurring objection: 5 of the 29 people filing complaints call the subscription too expensive, and some of those rate the app highly.66 posts / 7d
Get every post about Deepgram, filtered and labelled
This page is the summary. Tracking gives you the stream behind it for any brand, keyword or @mention: each new pain point, request and question as it’s posted, with the spam and promotions already taken out.