Speedrunning a modern title for 6-12 hours in a single session is already a physical feat. Adding high-quality live commentary on top of that, without dead air, voice fatigue, or keyboard clatter drowning out your callouts, is a separate discipline entirely. This guide covers the audio setup that lets you do both.
TL;DR
- Noise suppression removes keyboard and controller noise without a soundproof booth
- AI voice cloning preserves your commentary persona even when your actual voice is shot after hour 8
- low-latency audio capture routing into OBS adds under 15ms of audio latency — transparent during gameplay
- Calm, consistent delivery is more important than theatrical effects for speedrun commentary
- A comparison of common audio setups for speedrun streams is in the table below
Why Speedrun Streams Have Unique Audio Demands
Most streaming audio guides are written for casual gaming sessions — an hour or two, relaxed pace, mic in hand. Speedrunning inverts almost every assumption in those guides.
You are under time pressure, which means your voice is tense. You are doing the same segments dozens or hundreds of times across attempts, so your commentary needs to stay fresh even when you are not. Runs can span 6 to 12 hours, meaning voice fatigue is a real concern starting around hour four. And the mechanical input — fast keyboard sequences for PC games, rapid button mashing for console titles — creates continuous background noise that a standard mic setup does not handle well.
The speedrunning community has grown significantly as a streaming genre. Games like Super Mario 64, The Legend of Zelda: Ocarina of Time, Minecraft, and Dark Souls all have active speedrunning communities on Twitch and YouTube, and their top streamers average 4-8 hours per stream. The audio quality bar has risen accordingly — viewers in a 2026 speedrun stream expect the same production quality they would get from a podcast, not the muffled-keyboard ambience of early streaming.
Noise Suppression: The Most Important Tool You Are Not Using
Keyboard noise is the most common complaint in speedrun VOD reviews. A mechanical keyboard at full-speed input during a difficult segment produces a consistent 40-60 dB broadband noise floor that saturates around your voice signal. Dynamic mics reduce this — but only if you are within 5-10cm of the capsule, which is not practical during an active run.
Real-time noise suppression using a neural model trained on this specific category of noise removes it cleanly. The key difference from traditional noise gates is that a gate introduces silence artifacts — you hear the gate opening and closing during fast speech. Neural suppression operates continuously and preserves voice harmonics while removing the noise component, so your audio sounds like you are in a treated room even if you are not.
For speedrunning specifically, the relevant noise categories are:
- Mechanical keyboard (60WPM+ input during movement phases)
- Controller button noise (captures muffled through desk surface on hard mic mounting)
- Mouse clicks (relevant for PC-native titles like Minecraft Java, Celeste, Hollow Knight)
- Cooling fans (high-end PCs running at load produce consistent 200-600 Hz fan noise)
A good setup handles all four simultaneously with a single noise suppression pass.
Persona Consistency Across a 6-Hour Run
Speedrun commentary has a distinct persona challenge. The best speedrun commentators maintain a calm, analytical tone even during high-stakes late-game segments. Part of this is training — learning to separate emotional state from commentary delivery. But part of it is physical: a voice that starts naturally calm at hour one sounds strained and different by hour six.
Consistent delivery is what builds a loyal speedrun audience. Viewers who watch 3-4 hours into a VOD are there for your commentary as much as the run. If your voice changes character mid-stream — going from broadcast-quality clarity to raspy close-mic mumbling — it breaks the experience.
There are two practical approaches to managing this:
Approach 1: Compression and EQ as a guardrail. A gentle compressor set to 4:1 ratio with a -18 dBFS threshold smooths out the dynamic range between your fresh voice and your tired voice. A high-pass filter at 80 Hz removes the proximity-effect bass buildup that comes when you unconsciously lean closer to the mic as you get tired. This approach preserves your natural voice while making it more consistent.
Approach 2: AI voice cloning as a fallback. This is the more aggressive option and the one that more speedrunners are adopting. You record 10-30 minutes of clean commentary during your best vocal state — after warming up, before fatigue sets in. You train a personal AI clone from that recording. When your actual voice starts to show fatigue mid-stream, you activate the clone. Viewers hear your voice at its best throughout the run, not a degraded version of it.
The clone approach is not about misrepresenting yourself — it is the audio equivalent of color correction in video: preserving the intent of the original rather than broadcasting the artifact.
AI Cloning During Marathon Attempts
Marathon speedruns — defined loosely as any run where you are going for a personal best over multiple hours — have a specific pattern where AI cloning is most useful.
The first 90 minutes of most runs involve early-game segments you have completed hundreds of times. Commentary during these segments tends to be either absent (you are focused on execution) or repetitive. This is the ideal phase to use a clone — you can narrate what is happening without straining your voice before the segments that actually matter for the run.
Late-game segments, where a PB is within reach, demand the most from your commentary. Your voice is most strained precisely when the content is most interesting to viewers. Activating a pre-recorded-quality clone during high-pressure segments lets you focus entirely on execution while maintaining commentary presence.
The technical requirement for this approach is low end-to-end latency. You cannot have a 400ms delay between speaking and the audience hearing your voice — it disrupts your own natural speech rhythm and creates an uncanny valley effect where your mouth movements visible on webcam are out of sync with audio. Sub-300ms total processing time is the practical floor for real-time use; models operating at 80-150ms on dedicated hardware are comfortable for live streaming.
Setting Up low-latency audio capture Routing into OBS
The audio signal chain for a speedrun streaming setup is: microphone → voice changer (noise suppression + optional effects) → virtual output device → OBS audio input capture.
low-latency audio capture (Windows Audio Session API) is the Windows low-latency audio API that operates at the OS level. Voice changers using low-latency audio capture intercept your microphone signal before it reaches any other application, transform it, and output it to a virtual device. OBS then reads from that virtual device exactly as it would from a physical microphone.
The practical steps:
- In your voice changer software, set your physical microphone as the input and confirm the virtual output device name.
- In OBS Studio, go to Settings → Audio and set your microphone/auxiliary audio to the virtual output device from step 1.
- Add an Audio Input Capture source to your scene and confirm it is reading from the correct device.
- Open OBS’s Audio Mixer, right-click the microphone channel, and select Advanced Audio Properties. Set the sync offset to 0ms (the low-latency audio capture pipeline itself handles timing).
- Test with OBS’s built-in audio monitoring before going live — listen for latency, clipping, or noise suppression artifacts.
The entire signal chain from low-latency audio capture-based processing adds 10-15ms of audio latency. For reference, OBS’s own audio encoding adds another 20-40ms. The combined total is well under the 100ms threshold where audio-video sync becomes visible.
Which Games Benefit Most from This Setup
Super Mario 64 and Mario Category Runs
Mario speedruns are long even at world-record pace — any% SM64 is around 1:38 for current world record, but sub-record runs average 2-3 hours. Keyboard noise is not relevant for console emulation, but controller input and desk vibration are. The repetitive nature of early-game movement optimization makes commentary fatigue real. AI cloning shines here during Bowser fights — the same execution commentary repeated across 50+ attempts sounds identical with a clone active.
Minecraft Java Speedruns
Minecraft any% (random seed) is a PC-native title with heavy keyboard and mouse input. The current meta involves fast item crafting sequences, which produce very high keyboard noise. Noise suppression is arguably more important here than any voice effect. Runs are also unpredictable in length — a good seed can end in under 15 minutes, a bad one might take 45 — so per-session voice fatigue is less of an issue than per-attempt consistency.
The Legend of Zelda: Ocarina of Time
OoT speedruns are 17-20 minutes at elite level (Any% No IM/WW), but casual speedrunners attempting to break personal bests often stream 4-6 hours of attempts. The game’s long cutscenes and loading zones create natural low-commentary phases — exactly when clone activation makes sense. Many OoT runners develop a specific deadpan commentary style that a well-trained clone reproduces accurately.
Dark Souls and Elden Ring Runs
Souls speedruns have the most emotionally variable commentary of any category — calm analytical navigation punctuated by genuine emotional reactions to hits and deaths. Noise suppression for keyboard and mouse is high-priority given the precision input required. The emotional variability makes cloning less useful here than in other categories — viewers are watching specifically for authentic emotional reaction. Focus on clean suppression and compression rather than cloning for Souls runs.
Audio Setup Comparison for Speedrun Streamers
| Setup | Keyboard Noise | Voice Fatigue | OBS Latency | Setup Complexity |
|---|---|---|---|---|
| Dynamic mic, no processing | Poor | No help | ~5ms | Minimal |
| Dynamic mic + gate | Moderate | No help | ~5ms | Low |
| Condenser + noise suppression (software) | Good | No help | 10-20ms | Medium |
| Voice changer (DSP only) + low-latency audio capture | Good | Partial (compression) | 10-15ms | Medium |
| Voice changer (AI clone) + low-latency audio capture | Excellent | Full (clone covers fatigue) | 80-150ms | Medium-High |
The AI clone setup requires a one-time training investment of 20-40 minutes. After that, it is a single toggle during your stream setup.
Common Mistakes in Speedrun Audio Setup
Using a noise gate instead of noise suppression. Gates create abrupt silence artifacts when you pause between words — exactly the pattern of speedrun commentary, which involves a lot of short phrases and thinking pauses. Continuous neural suppression handles this without artifacts.
Setting the virtual audio device incorrectly in OBS. The most common cause of “my voice changer isn’t working in OBS” is OBS still reading from the physical microphone rather than the virtual output. Double-check both the Settings → Audio configuration and the individual scene’s audio capture source.
Applying OBS’s own noise suppression on top of software suppression. This causes double-processing artifacts — a metallic, hollow sound on voice harmonics. Use one or the other, not both.
Training an AI clone without adequate sample audio. A clone trained on 5 minutes of in-game mumbling will sound muddy. Train on 20-30 minutes of deliberate, clear commentary in the same acoustic environment you use for streaming.
Running AI processing on the same GPU as the game. On single-GPU systems, AI voice inference during a graphically intensive segment can cause brief frame drops. Use DSP-only processing during CPU-intensive or GPU-intensive game segments, and reserve AI cloning for lower-load phases.
The Broader Picture: Audio as a Competitive Differentiator
In a genre where run times are measured to the millisecond and improvement is incremental, the viewers who stick around for 6-hour attempts are specifically there for the commentary experience. Audio quality — or lack of it — is immediately apparent and immediately affects whether someone stays or leaves.
The speedrunners who built large followings on Twitch in the 2020s invested in their audio setups early. The barrier to entry for broadcast-quality audio has dropped significantly: the combination of noise suppression, smart compression, and AI voice tools means a single-person setup in a non-treated room can now produce audio that would have required a professional recording space five years ago.
The setup described in this guide requires no soundproofing, no hardware mixer, no external DSP unit, and no per-session configuration changes. Once it is running, your only job is the run.
FAQ
See the frontmatter FAQ section above for answers to common questions about latency, anti-cheat compatibility, noise suppression, OBS routing, and AI cloning for speedrun streams.