Local AI by VRAM Tier - 8GB | 12GB | 16GB | 24GB | 32GB
Deep dives: Chat | Coding | Translation | Vision | Speech | Agents
ElevenLabs charges $22/month for voice cloning. Otter.ai charges $17/month for transcription. Both send your audio to their servers. Running speech models locally gives you unlimited voice cloning, unlimited transcription, and complete privacy - all for the one-time cost of the GPU you already own.
Updated August 2026. Refreshed monthly as models and runtimes change.
This guide covers both halves of local speech: speech-to-text (STT / transcription) and text-to-speech (TTS / voice generation). The two tasks have very different hardware requirements, and you can often run both simultaneously.
What Changed in This Update
Four things, two of which are corrections to what this page said in July.
- Correction: the two top open-weight TTS models were listed in the wrong order. On the Artificial Analysis open-weights TTS board pulled 2026-08-02, Fish Audio S2 Pro is first and Step-Audio-EditX is second. The July version of this page had it the other way round. The recommendation below does not change, and the reason it does not change is licensing rather than ranking.
- Correction: whisper.cpp is at v1.9.2, not v1.8.6. The July version quoted a stale release. v1.9.0 also added native NVIDIA Parakeet support, which matters more than the version bump itself.
- A ranking claim has been withdrawn. The July version told you that the newer transcription models beat Whisper on English word error rate, sourced from the Open ASR Leaderboard. That ordering could not be reconfirmed for this update, so it has been removed rather than left standing. Details in Part 1.
- One new transcription model added: MOSS-Transcribe-Diarize 0.9B, which does speaker diarization alongside the transcript.
Part 1: Speech-to-Text (Transcription)
A Note on Rankings Before the Table
For three years, Whisper was the default answer for local transcription. It still supports 99+ languages and remains the broadest option, and a set of newer open-weight models now compete with it on English and major European languages.
What this page will no longer do is tell you the order they finish in. The July version of this guide published average word error rates taken from the Open ASR Leaderboard and summarised them as a podium. Re-checking that ranking for this update produced two mutually contradictory tables, one of which appeared to contain model names that do not exist, so none of it could be confirmed. The individual models below are all still current and were re-verified for this update; the live ranking between them was not. The leaderboard is linked above and is the place to check it yourself, on the day you read this.
That is a real downgrade in what this page can tell you, and it is the honest outcome. A stale podium left in place would have been worse.
The Models
| Model | Params | VRAM / Compute | License | Source |
|---|---|---|---|---|
| IBM Granite Speech 4.1 2B | 2B | Runs on a single consumer GPU | Apache 2.0 | HF card |
| NVIDIA Canary-Qwen-2.5B | 2.5B | ~A6000/RTX 5090 class | CC-BY-4.0 | HF card |
| NVIDIA Parakeet TDT 0.6B v3 | 600M | ~2 GB to load | CC-BY-4.0 | HF card |
| Whisper large-v3-turbo | 809M | ~6 GB | MIT | HF card |
| Whisper large-v3 | 1.5B | ~10 GB | Apache 2.0 | HF card |
Licences here are all permissive, but not identically so: Apache 2.0 and MIT let you ship a product, while CC-BY-4.0 obliges you to attribute. Read the card before a commercial deployment, and note that the two Whisper builds differ: large-v3 is Apache 2.0 on Hugging Face, large-v3-turbo is MIT.
New: MOSS-Transcribe-Diarize 0.9B
The one genuinely new entry since July is MOSS-Transcribe-Diarize 0.9B from the OpenMOSS-Team at Fudan University, released 2026-07-09 under Apache 2.0. It took first place at the INTERSPEECH 2026 MLC-SLM Challenge on 2026-07-14, and it has been downloaded more than 216,000 times.
Its distinguishing feature is scope. It produces the transcript, the speaker diarization and the timestamps in one pass, across 50+ languages, on audio up to 90 minutes long. If your input is meetings, interviews or podcasts, “who said this, and when” is usually as much of the job as “what was said,” and a single model that emits all three is a shorter pipeline than a transcriber plus a separate diarizer.
Two caveats. First, no VRAM figure is published for it, so it does not appear in the tier tables below; you will have to measure it on your own card. Second, the documented serving paths are vLLM and SGLang rather than the lighter single-binary runtimes, so plan for a server rather than a script.
Also worth knowing about, but not yet worth relying on: Hojo-ASR-V1 (HojoAI/Hojo-ASR-V1, Apache 2.0) was added to the Open ASR Leaderboard’s evaluation repository in late July. No repository named Hojo-ASR-4B exists; an earlier version of this page used that name in error. Its word error rate is self-reported and it has around 80 downloads, so treat it as an early entrant rather than an established option.
The Recommendation: Parakeet TDT 0.6B v3 for English, Whisper for everything else
Two picks, depending on your languages.
If you mostly transcribe English (or the 25 languages it covers), start with NVIDIA Parakeet TDT 0.6B v3. At 600M parameters it loads in roughly 2GB, runs extremely fast, and handles up to ~3 hours of audio in one pass with local attention. The tiny footprint is the whole argument: it leaves almost your entire GPU free for a TTS model or an LLM. Canary-Qwen-2.5B is the heavier English alternative, wanting more VRAM and trained on audio clips no longer than 40 seconds. Its card warns that longer sequences may technically run but with degraded accuracy, so plan on chunking rather than treating 40 seconds as a hard limit.
If you need broad multilingual coverage, stay on Whisper. Its 99+ language support is still unmatched by the newer models, and the tooling around it is excellent: whisper.cpp is at v1.9.2 (2026-08-04), and Faster-Whisper (CTranslate2) is roughly 4x faster on GPU than the reference code at the same accuracy. Whisper large-v3-turbo, at ~6GB, remains the sensible all-rounder for mixed-language audio.
The whisper.cpp release worth acting on is v1.9.0 (2026-06-17), which added native NVIDIA Parakeet support. That collapses a real annoyance: the English pick and the multilingual pick no longer need two separate stacks. Both can now run under the same C++ runtime, on the same GGML builds, instead of pairing whisper.cpp with a NeMo install for Parakeet. If you standardised on NeMo earlier in the year purely to get Parakeet, this is the update that lets you reconsider.
STT by VRAM Tier
| Tier | Best English pick | Best multilingual pick | VRAM used | Room for TTS? |
|---|---|---|---|---|
| 8GB | Parakeet TDT 0.6B v3 (~2GB) | Whisper large-v3-turbo (~6GB) | 2-6 GB | Yes - Parakeet leaves ~6GB free |
| 12GB | Parakeet / Canary-Qwen | Whisper large-v3-turbo | 2-6 GB | Yes - comfortable |
| 16GB | Canary-Qwen-2.5B | Whisper large-v3 (full) | 5-10 GB | Yes - most TTS models |
| 24GB | Canary-Qwen-2.5B | Whisper large-v3 (full) | 5-10 GB | Yes - everything fits |
| 32GB | Canary-Qwen-2.5B | Whisper large-v3 (full) | 5-10 GB | Yes - full pipeline + LLM |
MOSS-Transcribe-Diarize 0.9B is deliberately absent from this table: it publishes no VRAM figure, and inventing one would be worse than the gap.
The headline change from last year still holds: because Parakeet loads in ~2GB, the “STT eats most of my 8GB card” problem is largely gone. You can run transcription and a voice-cloning TTS model together on an entry GPU.
We have a detailed setup guide: How to Self-Host Whisper: Replace Otter.ai (the workflow is the same if you swap in Parakeet).
Part 2: Text-to-Speech (Voice Generation)
TTS is where the variety is, and where the ordering moved since July.
The Correction
The Artificial Analysis open-weights TTS leaderboard, pulled 2026-08-02:
| Model | Elo | License | VRAM |
|---|---|---|---|
| Fish Audio S2 Pro | 1121 | fish-audio-research-license (non-commercial) | ~17 GB (bf16, single-tester measurement) |
| Step-Audio-EditX | 1109 | Apache 2.0 (source repo) | ~12 GB (AWQ 4-bit ~6-8 GB) |
| Voxtral-4B-TTS-2603 | 1067 | CC-BY-NC-4.0 (non-commercial) | Not published |
| NVIDIA MagpieTTS Multilingual | 1065 | Contradictory, see below | Not published |
| Kokoro-82M | 1056 | Apache 2.0 | <1 GB, CPU-capable |
| Higgs Audio V3 | 1036 | Not checked in this pass | Not published |
| Chatterbox | 1014 | MIT | ~4-6 GB |
Fish Audio S2 Pro is now first and Step-Audio-EditX second. The July version of this page described Step-Audio-EditX as the top-ranked open TTS model. That was the order at the time and it is not the order now, so treat the July text as superseded rather than as a second opinion.
Two things this table is not. Elo is a leaderboard rating, not an accuracy measurement, so the roughly 12-point interval between the top two is not a statement about how good either one sounds on your text, in your language, with your reference voice. And several of these projects publish no VRAM number at all; where the column says “not published,” no figure was found for that model in this pass, and an invented estimate would be worse than the blank.
On MagpieTTS specifically: its licensing is internally contradictory. The model card names the NVIDIA Open Model License while the acknowledgement checkbox gating the download on Hugging Face states non-commercial use only. Those two cannot both be operative, and this guide is not going to guess which one a court would read. If you need it commercially, get that reconciled in writing by NVIDIA before you build on it.
The Recommendation: Step-Audio-EditX, and the reason is the licence
Step-Audio-EditX remains the pick even though it is no longer first on Elo, because Fish Audio S2 Pro ships under the fish-audio-research-license, which is non-commercial, while Step-Audio-EditX is Apache 2.0 in its source repository. One caveat worth knowing: its Hugging Face weights repo declares no licence field at all, so the Apache grant comes from the project repo rather than from the weights page.
For an audience that self-hosts, that distinction outranks an Elo gap of about 12 points. Apache 2.0 means you can put the output in a product, in a client deliverable, or in anything with a price on it, without asking anyone. A non-commercial licence means the first time your side project earns money you are either renegotiating with Fish Audio or re-recording everything. Rankings move every month. A licence is the part of the decision that follows you.
Fish Audio S2 Pro is still the better answer for personal projects and for the widest language coverage, at 80+ languages. Just read the licence file before it becomes load-bearing.
Model Breakdown
Step-Audio-EditX - StepFun’s 3B model, second on the August Elo pull and the highest-placed permissively licensed model among the entries in it. Beyond zero-shot cloning, its trick is editing: you can take generated or reference audio and iteratively push the emotion, style or delivery, rather than re-rolling from scratch. Apache 2.0. An AWQ 4-bit build brings it to roughly 6-8GB if you are tight on VRAM.
Fish Audio S2 Pro - The multilingual heavyweight, and first on the August Elo pull. S2 Pro covers 80+ languages. The catch, and it is a real one for a “run it yourself” audience: the weights ship under the non-commercial fish-audio-research-license. Fine for personal projects; a product needs a separate agreement with Fish Audio.
Kokoro-82M - The efficiency pick. At just 82M parameters, Kokoro runs on CPU and stays under 1GB, and it still places mid-table on the Elo board at 1055.49. It has no voice cloning (54 fixed voices across 8 languages), but for reading text aloud, notifications, or a local assistant’s output, the footprint is hard to argue with. Apache 2.0.
Chatterbox - The expressive cloner. Resemble AI’s Chatterbox does zero-shot cloning from a few seconds of audio, and its Multilingual V3 update pushed cross-language consistency and added emotion-exaggeration control across 23+ languages. MIT licensed, with built-in Perth watermarking on generated audio, at roughly 4-6GB.
Piper - The edge champion, under new management. The original rhasspy/piper repository is archived, with its last commit in August 2025, and its README now points only to OHF-Voice/piper1-gpl, where development continues. That matters for more than the URL: the maintained project is GPL-3.0, not MIT, so read it before embedding Piper in anything proprietary. Piper still runs on a Raspberry Pi at real-time speed with no GPU. Quality is a step below the neural cloners, but the near-zero footprint means it can run alongside anything, which suits home automation and smart-home responses.
Two older models are still in active use and have not been superseded for their niches: XTTS-v2 (Coqui, strong multilingual default) and F5-TTS (flow-matching zero-shot cloning). Neither appears among the entries in the August 2 Elo pull above, but plenty of existing pipelines still run them well.
TTS by VRAM Tier {#tts-tiers}
8GB VRAM {#8gb}
GPUs: RTX 4060, RTX 3060 8GB, RTX 3070
| Combo | STT | TTS | Total VRAM | Use Case |
|---|---|---|---|---|
| Speed focus | Parakeet TDT (2GB) | Kokoro-82M (<1GB) | ~3 GB | Fast transcription + readback, tons of headroom |
| Voice clone | Parakeet TDT (2GB) | Chatterbox (4-6GB) | ~6-8 GB | Transcribe and clone together |
| Edge/CPU | (CPU Whisper) | Piper (<1GB) | ~1 GB | 20+ languages, no GPU needed |
Best combo: Parakeet TDT 0.6B v3 + Chatterbox. A top-tier transcriber now costs ~2GB, so an 8GB card can run transcription and voice cloning simultaneously - something that required a 12GB+ card in the previous generation. Both are commercially usable (CC-BY-4.0 with attribution, and MIT).
12GB VRAM {#12gb}
GPUs: RTX 3060 12GB, RTX 4070
| Combo | STT | TTS | Total VRAM |
|---|---|---|---|
| Best quality | Parakeet TDT (2GB) | Step-Audio-EditX AWQ (6-8GB) | ~8-10 GB |
| Multilingual | Whisper turbo (6GB) | Chatterbox (4-6GB) | ~10-12 GB |
| Efficient | Parakeet TDT (2GB) | Kokoro (<1GB) | ~3 GB (+room for a small LLM) |
Best combo: Parakeet + Step-Audio-EditX (AWQ 4-bit). The best Apache 2.0 voice on the August board plus fast transcription, with room to spare.
16GB VRAM {#16gb}
GPUs: RTX 4060 Ti 16GB, RTX 5060, Arc A770
| Combo | STT | TTS | Total VRAM |
|---|---|---|---|
| Full quality | Canary-Qwen (5-6GB) | Step-Audio-EditX (12GB) | ~16-18 GB (tight - use AWQ build) |
| Voice assistant | Parakeet (2GB) | Chatterbox (6GB) | ~8 GB (+8GB for an LLM) |
| Multilingual | Whisper large-v3 (10GB) | Kokoro (<1GB) | ~11 GB |
Best combo: Parakeet + Chatterbox + a small LLM. 16GB is enough for a complete voice-in, AI-process, voice-out assistant with everything resident at once.
24GB VRAM {#24gb}
GPUs: RTX 3090, RTX 4090
| Combo | STT | TTS | LLM | Total VRAM |
|---|---|---|---|---|
| Voice assistant | Parakeet (2GB) | Step-Audio-EditX (12GB) | 8B-class (~6GB) | ~20 GB |
| Power pipeline | Canary-Qwen (6GB) | Fish S2 Pro (17GB) | - | ~23 GB (non-commercial licence) |
| Full stack | Parakeet (2GB) | Chatterbox (6GB) | 14B-class LLM (~11GB) | ~19 GB |
Best combo: Parakeet + Step-Audio-EditX + an 8B-class LLM. A complete local voice assistant - speak, get an answer read back in a cloned, emotion-controllable voice - with headroom left over, and nothing in the stack that blocks commercial use.
32GB VRAM {#32gb}
GPUs: RTX 5090
| Combo | STT | TTS | LLM | Total VRAM |
|---|---|---|---|---|
| Premium assistant | Canary-Qwen (6GB) | Step-Audio-EditX (12GB) | 14B-class LLM (~11GB) | ~29 GB |
| Multilingual studio | Whisper large-v3 (10GB) | Fish S2 Pro (17GB) | - | ~27 GB (non-commercial licence) |
Best combo: Canary-Qwen + Step-Audio-EditX + a 14B-class LLM. Transcription, expressive cloned speech and a capable conversational model, all resident on one card.
Cross-Tier Summary
Best STT by Tier
| Tier | English pick | Multilingual pick | Licence |
|---|---|---|---|
| 8GB+ | Parakeet TDT 0.6B v3 | Whisper large-v3-turbo | CC-BY-4.0 / MIT |
| 16GB+ | Canary-Qwen-2.5B | Whisper large-v3 | CC-BY-4.0 / Apache 2.0 |
| Any (server) | MOSS-Transcribe-Diarize 0.9B, when you need speaker labels too | MOSS-Transcribe-Diarize 0.9B, 50+ languages | Apache 2.0 |
No word error rates in this table on purpose. See the note at the top of Part 1.
Best TTS by Tier
| Tier | Model | Voice Clone | License | Notes |
|---|---|---|---|---|
| Any (CPU) | Piper / Kokoro-82M | No | GPL-3.0 / Apache 2.0 | Real-time, tiny footprint |
| 8-12GB | Chatterbox / Step-Audio-EditX AWQ | Yes | MIT / Apache 2.0 | Cloning on entry hardware |
| 16GB+ | Step-Audio-EditX | Yes | Apache 2.0 | Second on the August Elo pull, editable delivery, commercially usable |
| 16GB+ | Fish Audio S2 Pro | Yes | fish-audio-research-license | First on the August Elo pull, 80+ languages, non-commercial only |
Quick Start
Transcription with Faster-Whisper (multilingual)
pip install faster-whisper
# Python
from faster_whisper import WhisperModel
model = WhisperModel("large-v3-turbo", device="cuda")
segments, info = model.transcribe("meeting.mp3")
for segment in segments:
print(f"[{segment.start:.1f}s] {segment.text}")Transcription with Parakeet (English, via NVIDIA NeMo)
pip install nemo_toolkit['asr']
# Python
import nemo.collections.asr as nemo_asr
asr = nemo_asr.models.ASRModel.from_pretrained("nvidia/parakeet-tdt-0.6b-v3")
print(asr.transcribe(["meeting.wav"])[0].text)Since whisper.cpp v1.9.0 you can also run Parakeet under whisper.cpp itself, which is the lighter option if you are already using it for Whisper. Check the release notes for the current usage details.
TTS with Kokoro (zero GPU needed)
pip install kokoro-onnx
# Works on CPU - no GPU required
from kokoro_onnx import Kokoro
kokoro = Kokoro("kokoro-v1.0.onnx", "voices-v1.0.bin")
audio, sr = kokoro.create("Your text here", voice="af_heart")For complete setup guides with Docker, speaker identification, and production configurations:
All eleven local-AI guides, by hardware tier and by task, are indexed at Local AI by VRAM: Which Models Fit Your GPU.